Mixer/hermes-sync
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Sync all skills and memories 2026-04-14 07:27

Browse Source
This commit is contained in:
Mixer
2026-04-14 07:27:20 +09:00
parent 516bb44fe6
commit 1eba2bca95
386 changed files with 167655 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
skills/mlops/training/DESCRIPTION.md Normal file
View File

@@ -0,0 +1,3 @@
---
description: Fine-tuning, RLHF/DPO/GRPO training, distributed training frameworks, and optimization tools for training LLMs and other models.
---

161
skills/mlops/training/axolotl/SKILL.md Normal file
View File

@@ -0,0 +1,161 @@
---
name: axolotl
description: Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support
version: 1.0.0
author: Orchestra Research
license: MIT
dependencies: [axolotl, torch, transformers, datasets, peft, accelerate, deepspeed]
metadata:
hermes:
tags: [Fine-Tuning, Axolotl, LLM, LoRA, QLoRA, DPO, KTO, ORPO, GRPO, YAML, HuggingFace, DeepSpeed, Multimodal]
---
# Axolotl Skill
Comprehensive assistance with axolotl development, generated from official documentation.
## When to Use This Skill
This skill should be triggered when:
- Working with axolotl
- Asking about axolotl features or APIs
- Implementing axolotl solutions
- Debugging axolotl code
- Learning axolotl best practices
## Quick Reference
### Common Patterns
**Pattern 1:** To validate that acceptable data transfer speeds exist for your training job, running NCCL Tests can help pinpoint bottlenecks, for example:
```
./build/all_reduce_perf -b 8 -e 128M -f 2 -g 3
```
**Pattern 2:** Configure your model to use FSDP in the Axolotl yaml. For example:
```
fsdp_version: 2
fsdp_config:
offload_params: true
state_dict_type: FULL_STATE_DICT
auto_wrap_policy: TRANSFORMER_BASED_WRAP
transformer_layer_cls_to_wrap: LlamaDecoderLayer
reshard_after_forward: true
```
**Pattern 3:** The context_parallel_size should be a divisor of the total number of GPUs. For example:
```
context_parallel_size
```
**Pattern 4:** For example: - With 8 GPUs and no sequence parallelism: 8 different batches processed per step - With 8 GPUs and context_parallel_size=4: Only 2 different batches processed per step (each split across 4 GPUs) - If your per-GPU micro_batch_size is 2, the global batch size decreases from 16 to 4
```
context_parallel_size=4
```
**Pattern 5:** Setting save_compressed: true in your configuration enables saving models in a compressed format, which: - Reduces disk space usage by approximately 40% - Maintains compatibility with vLLM for accelerated inference - Maintains compatibility with llmcompressor for further optimization (example: quantization)
```
save_compressed: true
```
**Pattern 6:** Note It is not necessary to place your integration in the integrations folder. It can be in any location, so long as it’s installed in a package in your python env. See this repo for an example: https://github.com/axolotl-ai-cloud/diff-transformer
```
integrations
```
**Pattern 7:** Handle both single-example and batched data. - single example: sample[‘input_ids’] is a list[int] - batched data: sample[‘input_ids’] is a list[list[int]]
```
utils.trainer.drop_long_seq(sample, sequence_len=2048, min_sequence_len=2)
```
### Example Code Patterns
**Example 1** (python):
```python
cli.cloud.modal_.ModalCloud(config, app=None)
```
**Example 2** (python):
```python
cli.cloud.modal_.run_cmd(cmd, run_folder, volumes=None)
```
**Example 3** (python):
```python
core.trainers.base.AxolotlTrainer(
*_args,
bench_data_collator=None,
eval_data_collator=None,
dataset_tags=None,
**kwargs,
)
```
**Example 4** (python):
```python
core.trainers.base.AxolotlTrainer.log(logs, start_time=None)
```
**Example 5** (python):
```python
prompt_strategies.input_output.RawInputOutputPrompter()
```
## Reference Files
This skill includes comprehensive documentation in `references/`:
- **api.md** - Api documentation
- **dataset-formats.md** - Dataset-Formats documentation
- **other.md** - Other documentation
Use `view` to read specific reference files when detailed information is needed.
## Working with This Skill
### For Beginners
Start with the getting_started or tutorials reference files for foundational concepts.
### For Specific Features
Use the appropriate category reference file (api, guides, etc.) for detailed information.
### For Code Examples
The quick reference section above contains common patterns extracted from the official docs.
## Resources
### references/
Organized documentation extracted from official sources. These files contain:
- Detailed explanations
- Code examples with language annotations
- Links to original documentation
- Table of contents for quick navigation
### scripts/
Add helper scripts here for common automation tasks.
### assets/
Add templates, boilerplate, or example projects here.
## Notes
- This skill was automatically generated from official documentation
- Reference files preserve the structure and examples from source docs
- Code examples include language detection for better syntax highlighting
- Quick reference patterns are extracted from common usage examples in the docs
## Updating
To refresh this skill with updated documentation:
1. Re-run the scraper with the same configuration
2. The skill will be rebuilt with the latest information

5548
skills/mlops/training/axolotl/references/api.md Normal file
View File

File diff suppressed because it is too large Load Diff

1029
skills/mlops/training/axolotl/references/dataset-formats.md Normal file
View File

File diff suppressed because it is too large Load Diff

15
skills/mlops/training/axolotl/references/index.md Normal file
View File

@@ -0,0 +1,15 @@
# Axolotl Documentation Index
## Categories
### Api
**File:** `api.md`
**Pages:** 150
### Dataset-Formats
**File:** `dataset-formats.md`
**Pages:** 9
### Other
**File:** `other.md`
**Pages:** 26

3563
skills/mlops/training/axolotl/references/other.md Normal file
View File

File diff suppressed because it is too large Load Diff

97
skills/mlops/training/grpo-rl-training/README.md Normal file
View File

@@ -0,0 +1,97 @@
# GRPO/RL Training Skill
**Expert-level guidance for Group Relative Policy Optimization with TRL**
## 📁 Skill Structure
```
grpo-rl-training/
├── SKILL.md # Main skill documentation (READ THIS FIRST)
├── README.md # This file
├── templates/
│ └── basic_grpo_training.py # Production-ready training template
└── examples/
└── reward_functions_library.py # 20+ reward function examples
```
## 🚀 Quick Start
1. **Read SKILL.md** - Comprehensive guide with all concepts and patterns
2. **Copy `templates/basic_grpo_training.py`** - Start with working code
3. **Browse `examples/reward_functions_library.py`** - Pick reward functions for your task
4. **Modify for your use case** - Adapt dataset, rewards, and config
## 💡 What's Inside
### SKILL.md (Main Documentation)
- Core GRPO concepts and algorithm fundamentals
- Complete implementation workflow (dataset → rewards → training → deployment)
- 10+ reward function examples with code
- Hyperparameter tuning guide
- Training insights (loss behavior, metrics, debugging)
- Troubleshooting guide
- Production best practices
### Templates
- **basic_grpo_training.py**: Minimal, production-ready training script
- Uses Qwen 2.5 1.5B Instruct
- 3 reward functions (format + correctness)
- LoRA for efficient training
- Fully documented and ready to run
### Examples
- **reward_functions_library.py**: 20+ battle-tested reward functions
- Correctness rewards (exact match, fuzzy match, numeric, code execution)
- Format rewards (XML, JSON, strict/soft)
- Length rewards (ideal length, min/max)
- Style rewards (reasoning quality, citations, repetition penalty)
- Combined rewards (multi-objective optimization)
- Preset collections for common tasks
## 📖 Usage for Agents
When this skill is loaded in your agent's context:
1. **Always read SKILL.md first** before implementing
2. **Start simple** - Use length-based reward to validate setup
3. **Build incrementally** - Add one reward function at a time
4. **Reference examples** - Copy patterns from reward_functions_library.py
5. **Monitor training** - Watch reward metrics (not loss!)
## 🎯 Common Use Cases
| Task Type | Recommended Rewards | Template |
|-----------|---------------------|----------|
| Math reasoning | `MATH_REASONING_REWARDS` preset | basic_grpo_training.py |
| Code generation | `CODE_GENERATION_REWARDS` preset | Modify dataset in template |
| Summarization | `SUMMARIZATION_REWARDS` preset | Adjust prompts + rewards |
| Q&A | `QA_REWARDS` preset | Use fuzzy match + citations |
## ⚠️ Critical Reminders
- **Loss goes UP during training** - This is normal (it's KL divergence)
- **Use 3-5 reward functions** - Single rewards often fail
- **Test rewards before training** - Debug each function independently
- **Monitor reward_std** - Should stay > 0.1 (avoid mode collapse)
- **Start with num_generations=4-8** - Scale up if GPU allows
## 🔗 External Resources
- [TRL Documentation](https://huggingface.co/docs/trl)
- [DeepSeek R1 Paper](https://arxiv.org/abs/2501.12948)
- [Open R1 Implementation](https://github.com/huggingface/open-r1)
- [Unsloth (2-3x faster)](https://docs.unsloth.ai/)
## 📝 Version
**v1.0.0** - Initial release (January 2025)
## 👨‍💻 Maintained By
Orchestra Research
For questions or improvements, see https://orchestra.com
---
**License:** MIT
**Last Updated:** January 2025

575
skills/mlops/training/grpo-rl-training/SKILL.md Normal file
View File

@@ -0,0 +1,575 @@
---
name: grpo-rl-training
description: Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training
version: 1.0.0
author: Orchestra Research
license: MIT
dependencies: [transformers>=4.47.0, trl>=0.14.0, datasets>=3.2.0, peft>=0.14.0, torch]
metadata:
hermes:
tags: [Post-Training, Reinforcement Learning, GRPO, TRL, RLHF, Reward Modeling, Reasoning, DPO, PPO, Structured Output]
---
# GRPO/RL Training with TRL
Expert-level guidance for implementing Group Relative Policy Optimization (GRPO) using the Transformer Reinforcement Learning (TRL) library. This skill provides battle-tested patterns, critical insights, and production-ready workflows for fine-tuning language models with custom reward functions.
## When to Use This Skill
Use GRPO training when you need to:
- **Enforce specific output formats** (e.g., XML tags, JSON, structured reasoning)
- **Teach verifiable tasks** with objective correctness metrics (math, coding, fact-checking)
- **Improve reasoning capabilities** by rewarding chain-of-thought patterns
- **Align models to domain-specific behaviors** without labeled preference data
- **Optimize for multiple objectives** simultaneously (format + correctness + style)
**Do NOT use GRPO for:**
- Simple supervised fine-tuning tasks (use SFT instead)
- Tasks without clear reward signals
- When you already have high-quality preference pairs (use DPO/PPO instead)
---
## Core Concepts
### 1. GRPO Algorithm Fundamentals
**Key Mechanism:**
- Generates **multiple completions** for each prompt (group size: 4-16)
- Compares completions within each group using reward functions
- Updates policy to favor higher-rewarded responses relative to the group
**Critical Difference from PPO:**
- No separate reward model needed
- More sample-efficient (learns from within-group comparisons)
- Simpler to implement and debug
**Mathematical Intuition:**
```
For each prompt p:
1. Generate N completions: {c₁, c₂, ..., cₙ}
2. Compute rewards: {r₁, r₂, ..., rₙ}
3. Learn to increase probability of high-reward completions
relative to low-reward ones in the same group
```
### 2. Reward Function Design Philosophy
**Golden Rules:**
1. **Compose multiple reward functions** - Each handles one aspect (format, correctness, style)
2. **Scale rewards appropriately** - Higher weight = stronger signal
3. **Use incremental rewards** - Partial credit for partial compliance
4. **Test rewards independently** - Debug each reward function in isolation
**Reward Function Types:**
| Type | Use Case | Example Weight |
|------|----------|----------------|
| **Correctness** | Verifiable tasks (math, code) | 2.0 (highest) |
| **Format** | Strict structure enforcement | 0.5-1.0 |
| **Length** | Encourage verbosity/conciseness | 0.1-0.5 |
| **Style** | Penalize unwanted patterns | -0.5 to 0.5 |
---
## Implementation Workflow
### Step 1: Dataset Preparation
**Critical Requirements:**
- Prompts in chat format (list of dicts with 'role' and 'content')
- Include system prompts to set expectations
- For verifiable tasks, include ground truth answers as additional columns
**Example Structure:**
```python
from datasets import load_dataset, Dataset
SYSTEM_PROMPT = """
Respond in the following format:
<reasoning>
[Your step-by-step thinking]
</reasoning>
<answer>
[Final answer]
</answer>
"""
def prepare_dataset(raw_data):
"""
Transform raw data into GRPO-compatible format.
Returns: Dataset with columns:
- 'prompt': List[Dict] with role/content (system + user messages)
- 'answer': str (ground truth, optional but recommended)
"""
return raw_data.map(lambda x: {
'prompt': [
{'role': 'system', 'content': SYSTEM_PROMPT},
{'role': 'user', 'content': x['question']}
],
'answer': extract_answer(x['raw_answer'])
})
```
**Pro Tips:**
- Use one-shot or few-shot examples in system prompt for complex formats
- Keep prompts concise (max_prompt_length: 256-512 tokens)
- Validate data quality before training (garbage in = garbage out)
### Step 2: Reward Function Implementation
**Template Structure:**
```python
def reward_function_name(
prompts, # List[List[Dict]]: Original prompts
completions, # List[List[Dict]]: Model generations
answer=None, # Optional: Ground truth from dataset
**kwargs # Additional dataset columns
) -> list[float]:
"""
Evaluate completions and return rewards.
Returns: List of floats (one per completion)
"""
# Extract completion text
responses = [comp[0]['content'] for comp in completions]
# Compute rewards
rewards = []
for response in responses:
score = compute_score(response)
rewards.append(score)
return rewards
```
**Example 1: Correctness Reward (Math/Coding)**
```python
def correctness_reward(prompts, completions, answer, **kwargs):
"""Reward correct answers with high score."""
responses = [comp[0]['content'] for comp in completions]
extracted = [extract_final_answer(r) for r in responses]
return [2.0 if ans == gt else 0.0
for ans, gt in zip(extracted, answer)]
```
**Example 2: Format Reward (Structured Output)**
```python
import re
def format_reward(completions, **kwargs):
"""Reward XML-like structured format."""
pattern = r'<reasoning>.*?</reasoning>\s*<answer>.*?</answer>'
responses = [comp[0]['content'] for comp in completions]
return [1.0 if re.search(pattern, r, re.DOTALL) else 0.0
for r in responses]
```
**Example 3: Incremental Format Reward (Partial Credit)**
```python
def incremental_format_reward(completions, **kwargs):
"""Award partial credit for format compliance."""
responses = [comp[0]['content'] for comp in completions]
rewards = []
for r in responses:
score = 0.0
if '<reasoning>' in r:
score += 0.25
if '</reasoning>' in r:
score += 0.25
if '<answer>' in r:
score += 0.25
if '</answer>' in r:
score += 0.25
# Penalize extra text after closing tag
if r.count('</answer>') == 1:
extra_text = r.split('</answer>')[-1].strip()
score -= len(extra_text) * 0.001
rewards.append(score)
return rewards
```
**Critical Insight:**
Combine 3-5 reward functions for robust training. Order matters less than diversity of signals.
### Step 3: Training Configuration
**Memory-Optimized Config (Small GPU)**
```python
from trl import GRPOConfig
training_args = GRPOConfig(
output_dir="outputs/grpo-model",
# Learning rate
learning_rate=5e-6, # Lower = more stable
adam_beta1=0.9,
adam_beta2=0.99,
weight_decay=0.1,
warmup_ratio=0.1,
lr_scheduler_type='cosine',
# Batch settings
per_device_train_batch_size=1,
gradient_accumulation_steps=4, # Effective batch = 4
# GRPO-specific
num_generations=8, # Group size: 8-16 recommended
max_prompt_length=256,
max_completion_length=512,
# Training duration
num_train_epochs=1,
max_steps=None, # Or set fixed steps (e.g., 500)
# Optimization
bf16=True, # Faster on A100/H100
optim="adamw_8bit", # Memory-efficient optimizer
max_grad_norm=0.1,
# Logging
logging_steps=1,
save_steps=100,
report_to="wandb", # Or "none" for no logging
)
```
**High-Performance Config (Large GPU)**
```python
training_args = GRPOConfig(
output_dir="outputs/grpo-model",
learning_rate=1e-5,
per_device_train_batch_size=4,
gradient_accumulation_steps=2,
num_generations=16, # Larger groups = better signal
max_prompt_length=512,
max_completion_length=1024,
num_train_epochs=1,
bf16=True,
use_vllm=True, # Fast generation with vLLM
logging_steps=10,
)
```
**Critical Hyperparameters:**
| Parameter | Impact | Tuning Advice |
|-----------|--------|---------------|
| `num_generations` | Group size for comparison | Start with 8, increase to 16 if GPU allows |
| `learning_rate` | Convergence speed/stability | 5e-6 (safe), 1e-5 (faster, riskier) |
| `max_completion_length` | Output verbosity | Match your task (512 for reasoning, 256 for short answers) |
| `gradient_accumulation_steps` | Effective batch size | Increase if GPU memory limited |
### Step 4: Model Setup and Training
**Standard Setup (Transformers)**
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig
from trl import GRPOTrainer
# Load model
model_name = "Qwen/Qwen2.5-1.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2", # 2-3x faster
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
# Optional: LoRA for parameter-efficient training
peft_config = LoraConfig(
r=16, # Rank (higher = more capacity)
lora_alpha=32, # Scaling factor (typically 2*r)
target_modules=[
"q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj"
],
task_type="CAUSAL_LM",
lora_dropout=0.05,
)
# Initialize trainer
trainer = GRPOTrainer(
model=model,
processing_class=tokenizer,
reward_funcs=[
incremental_format_reward,
format_reward,
correctness_reward,
],
args=training_args,
train_dataset=dataset,
peft_config=peft_config, # Remove for full fine-tuning
)
# Train
trainer.train()
# Save
trainer.save_model("final_model")
```
**Unsloth Setup (2-3x Faster)**
```python
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="google/gemma-3-1b-it",
max_seq_length=1024,
load_in_4bit=True,
fast_inference=True,
max_lora_rank=32,
)
model = FastLanguageModel.get_peft_model(
model,
r=32,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj"],
lora_alpha=32,
use_gradient_checkpointing="unsloth",
)
# Rest is identical to standard setup
trainer = GRPOTrainer(model=model, ...)
trainer.train()
```
---
## Critical Training Insights
### 1. Loss Behavior (EXPECTED PATTERN)
- **Loss starts near 0 and INCREASES during training**
- This is CORRECT - loss measures KL divergence from initial policy
- Model is learning (diverging from original behavior to optimize rewards)
- Monitor reward metrics instead of loss for progress
### 2. Reward Tracking
Key metrics to watch:
- `reward`: Average across all completions
- `reward_std`: Diversity within groups (should remain > 0)
- `kl`: KL divergence from reference (should grow moderately)
**Healthy Training Pattern:**
```
Step Reward Reward_Std KL
100 0.5 0.3 0.02
200 0.8 0.25 0.05
300 1.2 0.2 0.08 ← Good progression
400 1.5 0.15 0.12
```
**Warning Signs:**
- Reward std → 0 (model collapsing to single response)
- KL exploding (> 0.5) (diverging too much, reduce LR)
- Reward stuck (reward functions too harsh or model capacity issue)
### 3. Common Pitfalls and Solutions
| Problem | Symptom | Solution |
|---------|---------|----------|
| **Mode collapse** | All completions identical | Increase `num_generations`, add diversity penalty |
| **No learning** | Flat rewards | Check reward function logic, increase LR |
| **OOM errors** | GPU memory exceeded | Reduce `num_generations`, enable gradient checkpointing |
| **Slow training** | < 1 it/s | Enable `use_vllm=True`, use Unsloth, reduce seq length |
| **Format ignored** | Model doesn't follow structure | Increase format reward weight, add incremental rewards |
---
## Advanced Patterns
### 1. Multi-Stage Training
For complex tasks, train in stages:
```python
# Stage 1: Format compliance (epochs=1)
trainer_stage1 = GRPOTrainer(
model=model,
reward_funcs=[incremental_format_reward, format_reward],
...
)
trainer_stage1.train()
# Stage 2: Correctness (epochs=1)
trainer_stage2 = GRPOTrainer(
model=model,
reward_funcs=[format_reward, correctness_reward],
...
)
trainer_stage2.train()
```
### 2. Adaptive Reward Scaling
```python
class AdaptiveReward:
def __init__(self, base_reward_func, initial_weight=1.0):
self.func = base_reward_func
self.weight = initial_weight
def __call__(self, *args, **kwargs):
rewards = self.func(*args, **kwargs)
return [r * self.weight for r in rewards]
def adjust_weight(self, success_rate):
"""Increase weight if model struggling, decrease if succeeding."""
if success_rate < 0.3:
self.weight *= 1.2
elif success_rate > 0.8:
self.weight *= 0.9
```
### 3. Custom Dataset Integration
```python
def load_custom_knowledge_base(csv_path):
"""Example: School communication platform docs."""
import pandas as pd
df = pd.read_csv(csv_path)
dataset = Dataset.from_pandas(df).map(lambda x: {
'prompt': [
{'role': 'system', 'content': CUSTOM_SYSTEM_PROMPT},
{'role': 'user', 'content': x['question']}
],
'answer': x['expert_answer']
})
return dataset
```
---
## Deployment and Inference
### Save and Merge LoRA
```python
# Merge LoRA adapters into base model
if hasattr(trainer.model, 'merge_and_unload'):
merged_model = trainer.model.merge_and_unload()
merged_model.save_pretrained("production_model")
tokenizer.save_pretrained("production_model")
```
### Inference Example
```python
from transformers import pipeline
generator = pipeline(
"text-generation",
model="production_model",
tokenizer=tokenizer
)
result = generator(
[
{'role': 'system', 'content': SYSTEM_PROMPT},
{'role': 'user', 'content': "What is 15 + 27?"}
],
max_new_tokens=256,
do_sample=True,
temperature=0.7,
top_p=0.9
)
print(result[0]['generated_text'])
```
---
## Best Practices Checklist
**Before Training:**
- [ ] Validate dataset format (prompts as List[Dict])
- [ ] Test reward functions on sample data
- [ ] Calculate expected max_prompt_length from data
- [ ] Choose appropriate num_generations based on GPU memory
- [ ] Set up logging (wandb recommended)
**During Training:**
- [ ] Monitor reward progression (should increase)
- [ ] Check reward_std (should stay > 0.1)
- [ ] Watch for OOM errors (reduce batch size if needed)
- [ ] Sample generations every 50-100 steps
- [ ] Validate format compliance on holdout set
**After Training:**
- [ ] Merge LoRA weights if using PEFT
- [ ] Test on diverse prompts
- [ ] Compare to baseline model
- [ ] Document reward weights and hyperparameters
- [ ] Save reproducibility config
---
## Troubleshooting Guide
### Debugging Workflow
1. **Isolate reward functions** - Test each independently
2. **Check data distribution** - Ensure diversity in prompts
3. **Reduce complexity** - Start with single reward, add gradually
4. **Monitor generations** - Print samples every N steps
5. **Validate extraction logic** - Ensure answer parsing works
### Quick Fixes
```python
# Debug reward function
def debug_reward(completions, **kwargs):
responses = [comp[0]['content'] for comp in completions]
for i, r in enumerate(responses[:2]): # Print first 2
print(f"Response {i}: {r[:200]}...")
return [1.0] * len(responses) # Dummy rewards
# Test without training
trainer = GRPOTrainer(..., reward_funcs=[debug_reward])
trainer.generate_completions(dataset[:1]) # Generate without updating
```
---
## References and Resources
**Official Documentation:**
- TRL GRPO Trainer: https://huggingface.co/docs/trl/grpo_trainer
- DeepSeek R1 Paper: https://arxiv.org/abs/2501.12948
- Unsloth Docs: https://docs.unsloth.ai/
**Example Repositories:**
- Open R1 Implementation: https://github.com/huggingface/open-r1
- TRL Examples: https://github.com/huggingface/trl/tree/main/examples
**Recommended Reading:**
- Progressive Disclosure Pattern for agent instructions
- Reward shaping in RL (Ng et al.)
- LoRA paper (Hu et al., 2021)
---
## Usage Instructions for Agents
When this skill is loaded:
1. **Read this entire file** before implementing GRPO training
2. **Start with the simplest reward function** (e.g., length-based) to validate setup
3. **Use the templates** in `templates/` directory as starting points
4. **Reference examples** in `examples/` for task-specific implementations
5. **Follow the workflow** sequentially (don't skip steps)
6. **Debug incrementally** - add one reward function at a time
**Critical Reminders:**
- Always use multiple reward functions (3-5 is optimal)
- Monitor reward metrics, not loss
- Test reward functions before training
- Start small (num_generations=4), scale up gradually
- Save checkpoints frequently (every 100 steps)
This skill is designed for **expert-level implementation**. Beginners should start with supervised fine-tuning before attempting GRPO.

228
skills/mlops/training/grpo-rl-training/templates/basic_grpo_training.py Normal file
View File

@@ -0,0 +1,228 @@
"""
Basic GRPO Training Template
=============================
A minimal, production-ready template for GRPO training with TRL.
Adapt this for your specific task by modifying:
1. Dataset loading (get_dataset function)
2. Reward functions (reward_*_func)
3. System prompt (SYSTEM_PROMPT)
4. Hyperparameters (GRPOConfig)
"""
import torch
import re
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig
from trl import GRPOTrainer, GRPOConfig
# ==================== CONFIGURATION ====================
MODEL_NAME = "Qwen/Qwen2.5-1.5B-Instruct"
OUTPUT_DIR = "outputs/grpo-model"
MAX_PROMPT_LENGTH = 256
MAX_COMPLETION_LENGTH = 512
SYSTEM_PROMPT = """
Respond in the following format:
<reasoning>
[Your step-by-step thinking]
</reasoning>
<answer>
[Final answer]
</answer>
"""
# ==================== DATASET ====================
def get_dataset(split="train"):
"""
Load and prepare your dataset.
Returns: Dataset with columns:
- 'prompt': List[Dict] with role/content
- 'answer': str (ground truth, optional)
"""
# Example: GSM8K math dataset
data = load_dataset('openai/gsm8k', 'main')[split]
def process_example(x):
# Extract ground truth answer
answer = x['answer'].split('####')[1].strip() if '####' in x['answer'] else None
return {
'prompt': [
{'role': 'system', 'content': SYSTEM_PROMPT},
{'role': 'user', 'content': x['question']}
],
'answer': answer
}
return data.map(process_example)
# ==================== HELPER FUNCTIONS ====================
def extract_xml_tag(text: str, tag: str) -> str:
"""Extract content between XML tags."""
pattern = f'<{tag}>(.*?)</{tag}>'
match = re.search(pattern, text, re.DOTALL)
return match.group(1).strip() if match else ""
def extract_answer(text: str) -> str:
"""Extract the final answer from structured output."""
return extract_xml_tag(text, 'answer')
# ==================== REWARD FUNCTIONS ====================
def correctness_reward_func(prompts, completions, answer, **kwargs):
"""
Reward correct answers.
Weight: 2.0 (highest priority)
"""
responses = [comp[0]['content'] for comp in completions]
extracted = [extract_answer(r) for r in responses]
return [2.0 if ans == gt else 0.0 for ans, gt in zip(extracted, answer)]
def format_reward_func(completions, **kwargs):
"""
Reward proper XML format.
Weight: 0.5
"""
pattern = r'<reasoning>.*?</reasoning>\s*<answer>.*?</answer>'
responses = [comp[0]['content'] for comp in completions]
return [0.5 if re.search(pattern, r, re.DOTALL) else 0.0 for r in responses]
def incremental_format_reward_func(completions, **kwargs):
"""
Incremental reward for partial format compliance.
Weight: up to 0.5
"""
responses = [comp[0]['content'] for comp in completions]
rewards = []
for r in responses:
score = 0.0
if '<reasoning>' in r:
score += 0.125
if '</reasoning>' in r:
score += 0.125
if '<answer>' in r:
score += 0.125
if '</answer>' in r:
score += 0.125
# Penalize extra content after closing tag
if '</answer>' in r:
extra = r.split('</answer>')[-1].strip()
score -= len(extra) * 0.001
rewards.append(score)
return rewards
# ==================== MODEL SETUP ====================
def setup_model_and_tokenizer():
"""Load model and tokenizer with optimizations."""
model = AutoModelForCausalLM.from_pretrained(
MODEL_NAME,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
tokenizer.pad_token = tokenizer.eos_token
return model, tokenizer
def get_peft_config():
"""LoRA configuration for parameter-efficient training."""
return LoraConfig(
r=16,
lora_alpha=32,
target_modules=[
"q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj"
],
task_type="CAUSAL_LM",
lora_dropout=0.05,
)
# ==================== TRAINING ====================
def main():
"""Main training function."""
# Load data
print("Loading dataset...")
dataset = get_dataset()
print(f"Dataset size: {len(dataset)}")
# Setup model
print("Loading model...")
model, tokenizer = setup_model_and_tokenizer()
# Training configuration
training_args = GRPOConfig(
output_dir=OUTPUT_DIR,
run_name="grpo-training",
# Learning rate
learning_rate=5e-6,
adam_beta1=0.9,
adam_beta2=0.99,
weight_decay=0.1,
warmup_ratio=0.1,
lr_scheduler_type='cosine',
# Batch settings
per_device_train_batch_size=1,
gradient_accumulation_steps=4,
# GRPO specific
num_generations=8,
max_prompt_length=MAX_PROMPT_LENGTH,
max_completion_length=MAX_COMPLETION_LENGTH,
# Training duration
num_train_epochs=1,
# Optimization
bf16=True,
optim="adamw_8bit",
max_grad_norm=0.1,
# Logging
logging_steps=1,
save_steps=100,
report_to="wandb", # Change to "none" to disable logging
)
# Initialize trainer
trainer = GRPOTrainer(
model=model,
processing_class=tokenizer,
reward_funcs=[
incremental_format_reward_func,
format_reward_func,
correctness_reward_func,
],
args=training_args,
train_dataset=dataset,
peft_config=get_peft_config(),
)
# Train
print("Starting training...")
trainer.train()
# Save final model
print(f"Saving model to {OUTPUT_DIR}/final")
trainer.save_model(f"{OUTPUT_DIR}/final")
print("Training complete!")
if __name__ == "__main__":
main()

434
skills/mlops/training/peft/SKILL.md Normal file
View File

@@ -0,0 +1,434 @@
---
name: peft-fine-tuning
description: Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. Use when fine-tuning large models (7B-70B) with limited GPU memory, when you need to train <1% of parameters with minimal accuracy loss, or for multi-adapter serving. HuggingFace's official library integrated with transformers ecosystem.
version: 1.0.0
author: Orchestra Research
license: MIT
dependencies: [peft>=0.13.0, transformers>=4.45.0, torch>=2.0.0, bitsandbytes>=0.43.0]
metadata:
hermes:
tags: [Fine-Tuning, PEFT, LoRA, QLoRA, Parameter-Efficient, Adapters, Low-Rank, Memory Optimization, Multi-Adapter]
---
# PEFT (Parameter-Efficient Fine-Tuning)
Fine-tune LLMs by training <1% of parameters using LoRA, QLoRA, and 25+ adapter methods.
## When to use PEFT
**Use PEFT/LoRA when:**
- Fine-tuning 7B-70B models on consumer GPUs (RTX 4090, A100)
- Need to train <1% parameters (6MB adapters vs 14GB full model)
- Want fast iteration with multiple task-specific adapters
- Deploying multiple fine-tuned variants from one base model
**Use QLoRA (PEFT + quantization) when:**
- Fine-tuning 70B models on single 24GB GPU
- Memory is the primary constraint
- Can accept ~5% quality trade-off vs full fine-tuning
**Use full fine-tuning instead when:**
- Training small models (<1B parameters)
- Need maximum quality and have compute budget
- Significant domain shift requires updating all weights
## Quick start
### Installation
```bash
# Basic installation
pip install peft
# With quantization support (recommended)
pip install peft bitsandbytes
# Full stack
pip install peft transformers accelerate bitsandbytes datasets
```
### LoRA fine-tuning (standard)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import get_peft_model, LoraConfig, TaskType
from datasets import load_dataset
# Load base model
model_name = "meta-llama/Llama-3.1-8B"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
# LoRA configuration
lora_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
r=16, # Rank (8-64, higher = more capacity)
lora_alpha=32, # Scaling factor (typically 2*r)
lora_dropout=0.05, # Dropout for regularization
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"], # Attention layers
bias="none" # Don't train biases
)
# Apply LoRA
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# Output: trainable params: 13,631,488 || all params: 8,043,307,008 || trainable%: 0.17%
# Prepare dataset
dataset = load_dataset("databricks/databricks-dolly-15k", split="train")
def tokenize(example):
text = f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['response']}"
return tokenizer(text, truncation=True, max_length=512, padding="max_length")
tokenized = dataset.map(tokenize, remove_columns=dataset.column_names)
# Training
training_args = TrainingArguments(
output_dir="./lora-llama",
num_train_epochs=3,
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
learning_rate=2e-4,
fp16=True,
logging_steps=10,
save_strategy="epoch"
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized,
data_collator=lambda data: {"input_ids": torch.stack([f["input_ids"] for f in data]),
"attention_mask": torch.stack([f["attention_mask"] for f in data]),
"labels": torch.stack([f["input_ids"] for f in data])}
)
trainer.train()
# Save adapter only (6MB vs 16GB)
model.save_pretrained("./lora-llama-adapter")
```
### QLoRA fine-tuning (memory-efficient)
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import get_peft_model, LoraConfig, prepare_model_for_kbit_training
# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4", # NormalFloat4 (best for LLMs)
bnb_4bit_compute_dtype="bfloat16", # Compute in bf16
bnb_4bit_use_double_quant=True # Nested quantization
)
# Load quantized model
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-3.1-70B",
quantization_config=bnb_config,
device_map="auto"
)
# Prepare for training (enables gradient checkpointing)
model = prepare_model_for_kbit_training(model)
# LoRA config for QLoRA
lora_config = LoraConfig(
r=64, # Higher rank for 70B
lora_alpha=128,
lora_dropout=0.1,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
bias="none",
task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
# 70B model now fits on single 24GB GPU!
```
## LoRA parameter selection
### Rank (r) - capacity vs efficiency
| Rank | Trainable Params | Memory | Quality | Use Case |
|------|-----------------|--------|---------|----------|
| 4 | ~3M | Minimal | Lower | Simple tasks, prototyping |
| **8** | ~7M | Low | Good | **Recommended starting point** |
| **16** | ~14M | Medium | Better | **General fine-tuning** |
| 32 | ~27M | Higher | High | Complex tasks |
| 64 | ~54M | High | Highest | Domain adaptation, 70B models |
### Alpha (lora_alpha) - scaling factor
```python
# Rule of thumb: alpha = 2 * rank
LoraConfig(r=16, lora_alpha=32) # Standard
LoraConfig(r=16, lora_alpha=16) # Conservative (lower learning rate effect)
LoraConfig(r=16, lora_alpha=64) # Aggressive (higher learning rate effect)
```
### Target modules by architecture
```python
# Llama / Mistral / Qwen
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
# GPT-2 / GPT-Neo
target_modules = ["c_attn", "c_proj", "c_fc"]
# Falcon
target_modules = ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"]
# BLOOM
target_modules = ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"]
# Auto-detect all linear layers
target_modules = "all-linear" # PEFT 0.6.0+
```
## Loading and merging adapters
### Load trained adapter
```python
from peft import PeftModel, AutoPeftModelForCausalLM
from transformers import AutoModelForCausalLM
# Option 1: Load with PeftModel
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
model = PeftModel.from_pretrained(base_model, "./lora-llama-adapter")
# Option 2: Load directly (recommended)
model = AutoPeftModelForCausalLM.from_pretrained(
"./lora-llama-adapter",
device_map="auto"
)
```
### Merge adapter into base model
```python
# Merge for deployment (no adapter overhead)
merged_model = model.merge_and_unload()
# Save merged model
merged_model.save_pretrained("./llama-merged")
tokenizer.save_pretrained("./llama-merged")
# Push to Hub
merged_model.push_to_hub("username/llama-finetuned")
```
### Multi-adapter serving
```python
from peft import PeftModel
# Load base with first adapter
model = AutoPeftModelForCausalLM.from_pretrained("./adapter-task1")
# Load additional adapters
model.load_adapter("./adapter-task2", adapter_name="task2")
model.load_adapter("./adapter-task3", adapter_name="task3")
# Switch between adapters at runtime
model.set_adapter("task1") # Use task1 adapter
output1 = model.generate(**inputs)
model.set_adapter("task2") # Switch to task2
output2 = model.generate(**inputs)
# Disable adapters (use base model)
with model.disable_adapter():
base_output = model.generate(**inputs)
```
## PEFT methods comparison
| Method | Trainable % | Memory | Speed | Best For |
|--------|------------|--------|-------|----------|
| **LoRA** | 0.1-1% | Low | Fast | General fine-tuning |
| **QLoRA** | 0.1-1% | Very Low | Medium | Memory-constrained |
| AdaLoRA | 0.1-1% | Low | Medium | Automatic rank selection |
| IA3 | 0.01% | Minimal | Fastest | Few-shot adaptation |
| Prefix Tuning | 0.1% | Low | Medium | Generation control |
| Prompt Tuning | 0.001% | Minimal | Fast | Simple task adaptation |
| P-Tuning v2 | 0.1% | Low | Medium | NLU tasks |
### IA3 (minimal parameters)
```python
from peft import IA3Config
ia3_config = IA3Config(
target_modules=["q_proj", "v_proj", "k_proj", "down_proj"],
feedforward_modules=["down_proj"]
)
model = get_peft_model(model, ia3_config)
# Trains only 0.01% of parameters!
```
### Prefix Tuning
```python
from peft import PrefixTuningConfig
prefix_config = PrefixTuningConfig(
task_type="CAUSAL_LM",
num_virtual_tokens=20, # Prepended tokens
prefix_projection=True # Use MLP projection
)
model = get_peft_model(model, prefix_config)
```
## Integration patterns
### With TRL (SFTTrainer)
```python
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig
lora_config = LoraConfig(r=16, lora_alpha=32, target_modules="all-linear")
trainer = SFTTrainer(
model=model,
args=SFTConfig(output_dir="./output", max_seq_length=512),
train_dataset=dataset,
peft_config=lora_config, # Pass LoRA config directly
)
trainer.train()
```
### With Axolotl (YAML config)
```yaml
# axolotl config.yaml
adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
- q_proj
- v_proj
- k_proj
- o_proj
lora_target_linear: true # Target all linear layers
```
### With vLLM (inference)
```python
from vllm import LLM
from vllm.lora.request import LoRARequest
# Load base model with LoRA support
llm = LLM(model="meta-llama/Llama-3.1-8B", enable_lora=True)
# Serve with adapter
outputs = llm.generate(
prompts,
lora_request=LoRARequest("adapter1", 1, "./lora-adapter")
)
```
## Performance benchmarks
### Memory usage (Llama 3.1 8B)
| Method | GPU Memory | Trainable Params |
|--------|-----------|------------------|
| Full fine-tuning | 60+ GB | 8B (100%) |
| LoRA r=16 | 18 GB | 14M (0.17%) |
| QLoRA r=16 | 6 GB | 14M (0.17%) |
| IA3 | 16 GB | 800K (0.01%) |
### Training speed (A100 80GB)
| Method | Tokens/sec | vs Full FT |
|--------|-----------|------------|
| Full FT | 2,500 | 1x |
| LoRA | 3,200 | 1.3x |
| QLoRA | 2,100 | 0.84x |
### Quality (MMLU benchmark)
| Model | Full FT | LoRA | QLoRA |
|-------|---------|------|-------|
| Llama 2-7B | 45.3 | 44.8 | 44.1 |
| Llama 2-13B | 54.8 | 54.2 | 53.5 |
## Common issues
### CUDA OOM during training
```python
# Solution 1: Enable gradient checkpointing
model.gradient_checkpointing_enable()
# Solution 2: Reduce batch size + increase accumulation
TrainingArguments(
per_device_train_batch_size=1,
gradient_accumulation_steps=16
)
# Solution 3: Use QLoRA
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
```
### Adapter not applying
```python
# Verify adapter is active
print(model.active_adapters) # Should show adapter name
# Check trainable parameters
model.print_trainable_parameters()
# Ensure model in training mode
model.train()
```
### Quality degradation
```python
# Increase rank
LoraConfig(r=32, lora_alpha=64)
# Target more modules
target_modules = "all-linear"
# Use more training data and epochs
TrainingArguments(num_train_epochs=5)
# Lower learning rate
TrainingArguments(learning_rate=1e-4)
```
## Best practices
1. **Start with r=8-16**, increase if quality insufficient
2. **Use alpha = 2 * rank** as starting point
3. **Target attention + MLP layers** for best quality/efficiency
4. **Enable gradient checkpointing** for memory savings
5. **Save adapters frequently** (small files, easy rollback)
6. **Evaluate on held-out data** before merging
7. **Use QLoRA for 70B+ models** on consumer hardware
## References
- **[Advanced Usage](references/advanced-usage.md)** - DoRA, LoftQ, rank stabilization, custom modules
- **[Troubleshooting](references/troubleshooting.md)** - Common errors, debugging, optimization
## Resources
- **GitHub**: https://github.com/huggingface/peft
- **Docs**: https://huggingface.co/docs/peft
- **LoRA Paper**: arXiv:2106.09685
- **QLoRA Paper**: arXiv:2305.14314
- **Models**: https://huggingface.co/models?library=peft

514
skills/mlops/training/peft/references/advanced-usage.md Normal file
View File

@@ -0,0 +1,514 @@
# PEFT Advanced Usage Guide
## Advanced LoRA Variants
### DoRA (Weight-Decomposed Low-Rank Adaptation)
DoRA decomposes weights into magnitude and direction components, often achieving better results than standard LoRA:
```python
from peft import LoraConfig
dora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
use_dora=True, # Enable DoRA
task_type="CAUSAL_LM"
)
model = get_peft_model(model, dora_config)
```
**When to use DoRA**:
- Consistently outperforms LoRA on instruction-following tasks
- Slightly higher memory (~10%) due to magnitude vectors
- Best for quality-critical fine-tuning
### AdaLoRA (Adaptive Rank)
Automatically adjusts rank per layer based on importance:
```python
from peft import AdaLoraConfig
adalora_config = AdaLoraConfig(
init_r=64, # Initial rank
target_r=16, # Target average rank
tinit=200, # Warmup steps
tfinal=1000, # Final pruning step
deltaT=10, # Rank update frequency
beta1=0.85,
beta2=0.85,
orth_reg_weight=0.5, # Orthogonality regularization
target_modules=["q_proj", "v_proj"],
task_type="CAUSAL_LM"
)
```
**Benefits**:
- Allocates more rank to important layers
- Can reduce total parameters while maintaining quality
- Good for exploring optimal rank distribution
### LoRA+ (Asymmetric Learning Rates)
Different learning rates for A and B matrices:
```python
from peft import LoraConfig
# LoRA+ uses higher LR for B matrix
lora_plus_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules="all-linear",
use_rslora=True, # Rank-stabilized LoRA (related technique)
)
# Manual implementation of LoRA+
from torch.optim import AdamW
# Group parameters
lora_A_params = [p for n, p in model.named_parameters() if "lora_A" in n]
lora_B_params = [p for n, p in model.named_parameters() if "lora_B" in n]
optimizer = AdamW([
{"params": lora_A_params, "lr": 1e-4},
{"params": lora_B_params, "lr": 1e-3}, # 10x higher for B
])
```
### rsLoRA (Rank-Stabilized LoRA)
Scales LoRA outputs to stabilize training with different ranks:
```python
lora_config = LoraConfig(
r=64,
lora_alpha=64,
use_rslora=True, # Enables rank-stabilized scaling
target_modules="all-linear"
)
```
**When to use**:
- When experimenting with different ranks
- Helps maintain consistent behavior across rank values
- Recommended for r > 32
## LoftQ (LoRA-Fine-Tuning-aware Quantization)
Initializes LoRA weights to compensate for quantization error:
```python
from peft import LoftQConfig, LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
# LoftQ configuration
loftq_config = LoftQConfig(
loftq_bits=4, # Quantization bits
loftq_iter=5, # Alternating optimization iterations
)
# LoRA config with LoftQ initialization
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules="all-linear",
init_lora_weights="loftq",
loftq_config=loftq_config,
task_type="CAUSAL_LM"
)
# Load quantized model
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-3.1-8B",
quantization_config=bnb_config
)
model = get_peft_model(model, lora_config)
```
**Benefits over standard QLoRA**:
- Better initial quality after quantization
- Faster convergence
- ~1-2% better final accuracy on benchmarks
## Custom Module Targeting
### Target specific layers
```python
# Target only first and last transformer layers
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules=["model.layers.0.self_attn.q_proj",
"model.layers.0.self_attn.v_proj",
"model.layers.31.self_attn.q_proj",
"model.layers.31.self_attn.v_proj"],
layers_to_transform=[0, 31] # Alternative approach
)
```
### Layer pattern matching
```python
# Target layers 0-10 only
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules="all-linear",
layers_to_transform=list(range(11)), # Layers 0-10
layers_pattern="model.layers"
)
```
### Exclude specific layers
```python
lora_config = LoraConfig(
r=16,
target_modules="all-linear",
modules_to_save=["lm_head"], # Train these fully (not LoRA)
)
```
## Embedding and LM Head Training
### Train embeddings with LoRA
```python
from peft import LoraConfig
# Include embeddings
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules=["q_proj", "v_proj", "embed_tokens"], # Include embeddings
modules_to_save=["lm_head"], # Train lm_head fully
)
```
### Extending vocabulary with LoRA
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import get_peft_model, LoraConfig
# Add new tokens
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")
new_tokens = ["<custom_token_1>", "<custom_token_2>"]
tokenizer.add_tokens(new_tokens)
# Resize model embeddings
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
model.resize_token_embeddings(len(tokenizer))
# Configure LoRA to train new embeddings
lora_config = LoraConfig(
r=16,
target_modules="all-linear",
modules_to_save=["embed_tokens", "lm_head"], # Train these fully
)
model = get_peft_model(model, lora_config)
```
## Multi-Adapter Patterns
### Adapter composition
```python
from peft import PeftModel
# Load model with multiple adapters
model = AutoPeftModelForCausalLM.from_pretrained("./base-adapter")
model.load_adapter("./style-adapter", adapter_name="style")
model.load_adapter("./task-adapter", adapter_name="task")
# Combine adapters (weighted sum)
model.add_weighted_adapter(
adapters=["style", "task"],
weights=[0.7, 0.3],
adapter_name="combined",
combination_type="linear" # or "cat", "svd"
)
model.set_adapter("combined")
```
### Adapter stacking
```python
# Stack adapters (apply sequentially)
model.add_weighted_adapter(
adapters=["base", "domain", "task"],
weights=[1.0, 1.0, 1.0],
adapter_name="stacked",
combination_type="cat" # Concatenate adapter outputs
)
```
### Dynamic adapter switching
```python
import torch
class MultiAdapterModel:
def __init__(self, base_model_path, adapter_paths):
self.model = AutoPeftModelForCausalLM.from_pretrained(adapter_paths[0])
for name, path in adapter_paths[1:].items():
self.model.load_adapter(path, adapter_name=name)
def generate(self, prompt, adapter_name="default"):
self.model.set_adapter(adapter_name)
return self.model.generate(**self.tokenize(prompt))
def generate_ensemble(self, prompt, adapters, weights):
"""Generate with weighted adapter ensemble"""
outputs = []
for adapter, weight in zip(adapters, weights):
self.model.set_adapter(adapter)
logits = self.model(**self.tokenize(prompt)).logits
outputs.append(weight * logits)
return torch.stack(outputs).sum(dim=0)
```
## Memory Optimization
### Gradient checkpointing with LoRA
```python
from peft import prepare_model_for_kbit_training
# Enable gradient checkpointing
model = prepare_model_for_kbit_training(
model,
use_gradient_checkpointing=True,
gradient_checkpointing_kwargs={"use_reentrant": False}
)
```
### CPU offloading for training
```python
from accelerate import Accelerator
accelerator = Accelerator(
mixed_precision="bf16",
gradient_accumulation_steps=8,
cpu_offload=True # Offload optimizer states to CPU
)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
```
### Memory-efficient attention with LoRA
```python
from transformers import AutoModelForCausalLM
# Combine Flash Attention 2 with LoRA
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-3.1-8B",
attn_implementation="flash_attention_2",
torch_dtype=torch.bfloat16
)
# Apply LoRA
model = get_peft_model(model, lora_config)
```
## Inference Optimization
### Merge for deployment
```python
# Merge adapter weights into base model
merged_model = model.merge_and_unload()
# Quantize merged model for inference
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(load_in_4bit=True)
quantized_model = AutoModelForCausalLM.from_pretrained(
"./merged-model",
quantization_config=bnb_config
)
```
### Export to different formats
```python
# Export to GGUF (llama.cpp)
# First merge, then convert
merged_model.save_pretrained("./merged-model")
# Use llama.cpp converter
# python convert-hf-to-gguf.py ./merged-model --outfile model.gguf
# Export to ONNX
from optimum.onnxruntime import ORTModelForCausalLM
ort_model = ORTModelForCausalLM.from_pretrained(
"./merged-model",
export=True
)
ort_model.save_pretrained("./onnx-model")
```
### Batch adapter inference
```python
from vllm import LLM
from vllm.lora.request import LoRARequest
# Initialize with LoRA support
llm = LLM(
model="meta-llama/Llama-3.1-8B",
enable_lora=True,
max_lora_rank=64,
max_loras=4 # Max concurrent adapters
)
# Batch with different adapters
requests = [
("prompt1", LoRARequest("adapter1", 1, "./adapter1")),
("prompt2", LoRARequest("adapter2", 2, "./adapter2")),
("prompt3", LoRARequest("adapter1", 1, "./adapter1")),
]
outputs = llm.generate(
[r[0] for r in requests],
lora_request=[r[1] for r in requests]
)
```
## Training Recipes
### Instruction tuning recipe
```python
lora_config = LoraConfig(
r=16,
lora_alpha=32,
lora_dropout=0.05,
target_modules="all-linear",
bias="none",
task_type="CAUSAL_LM"
)
training_args = TrainingArguments(
output_dir="./output",
num_train_epochs=3,
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
learning_rate=2e-4,
lr_scheduler_type="cosine",
warmup_ratio=0.03,
bf16=True,
logging_steps=10,
save_strategy="steps",
save_steps=100,
eval_strategy="steps",
eval_steps=100,
)
```
### Code generation recipe
```python
lora_config = LoraConfig(
r=32, # Higher rank for code
lora_alpha=64,
lora_dropout=0.1,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
bias="none",
task_type="CAUSAL_LM"
)
training_args = TrainingArguments(
learning_rate=1e-4, # Lower LR for code
num_train_epochs=2,
max_seq_length=2048, # Longer sequences
)
```
### Conversational/Chat recipe
```python
from trl import SFTTrainer
lora_config = LoraConfig(
r=16,
lora_alpha=16, # alpha = r for chat
lora_dropout=0.05,
target_modules="all-linear"
)
# Use chat template
def format_chat(example):
messages = [
{"role": "user", "content": example["instruction"]},
{"role": "assistant", "content": example["response"]}
]
return tokenizer.apply_chat_template(messages, tokenize=False)
trainer = SFTTrainer(
model=model,
peft_config=lora_config,
train_dataset=dataset.map(format_chat),
max_seq_length=1024,
)
```
## Debugging and Validation
### Verify adapter application
```python
# Check which modules have LoRA
for name, module in model.named_modules():
if hasattr(module, "lora_A"):
print(f"LoRA applied to: {name}")
# Print detailed config
print(model.peft_config)
# Check adapter state
print(f"Active adapters: {model.active_adapters}")
print(f"Trainable: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
```
### Compare with base model
```python
# Generate with adapter
model.set_adapter("default")
adapter_output = model.generate(**inputs)
# Generate without adapter
with model.disable_adapter():
base_output = model.generate(**inputs)
print(f"Adapter: {tokenizer.decode(adapter_output[0])}")
print(f"Base: {tokenizer.decode(base_output[0])}")
```
### Monitor training metrics
```python
from transformers import TrainerCallback
class LoRACallback(TrainerCallback):
def on_log(self, args, state, control, logs=None, **kwargs):
if "loss" in logs:
# Log adapter-specific metrics
model = kwargs["model"]
lora_params = sum(p.numel() for n, p in model.named_parameters()
if "lora" in n and p.requires_grad)
print(f"Step {state.global_step}: loss={logs['loss']:.4f}, lora_params={lora_params}")
```

480
skills/mlops/training/peft/references/troubleshooting.md Normal file
View File

@@ -0,0 +1,480 @@
# PEFT Troubleshooting Guide
## Installation Issues
### bitsandbytes CUDA Error
**Error**: `CUDA Setup failed despite GPU being available`
**Fix**:
```bash
# Check CUDA version
nvcc --version
# Install matching bitsandbytes
pip uninstall bitsandbytes
pip install bitsandbytes --no-cache-dir
# Or compile from source for specific CUDA
git clone https://github.com/TimDettmers/bitsandbytes.git
cd bitsandbytes
CUDA_VERSION=118 make cuda11x # Adjust for your CUDA
pip install .
```
### Triton Import Error
**Error**: `ModuleNotFoundError: No module named 'triton'`
**Fix**:
```bash
# Install triton (Linux only)
pip install triton
# Windows: Triton not supported, use CUDA backend
# Set environment variable to disable triton
export CUDA_VISIBLE_DEVICES=0
```
### PEFT Version Conflicts
**Error**: `AttributeError: 'LoraConfig' object has no attribute 'use_dora'`
**Fix**:
```bash
# Upgrade to latest PEFT
pip install peft>=0.13.0 --upgrade
# Check version
python -c "import peft; print(peft.__version__)"
```
## Training Issues
### CUDA Out of Memory
**Error**: `torch.cuda.OutOfMemoryError: CUDA out of memory`
**Solutions**:
1. **Enable gradient checkpointing**:
```python
from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)
```
2. **Reduce batch size**:
```python
TrainingArguments(
per_device_train_batch_size=1,
gradient_accumulation_steps=16 # Maintain effective batch size
)
```
3. **Use QLoRA**:
```python
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True
)
model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=bnb_config)
```
4. **Lower LoRA rank**:
```python
LoraConfig(r=8) # Instead of r=16 or higher
```
5. **Target fewer modules**:
```python
target_modules=["q_proj", "v_proj"] # Instead of all-linear
```
### Loss Not Decreasing
**Problem**: Training loss stays flat or increases.
**Solutions**:
1. **Check learning rate**:
```python
# Start lower
TrainingArguments(learning_rate=1e-4) # Not 2e-4 or higher
```
2. **Verify adapter is active**:
```python
model.print_trainable_parameters()
# Should show >0 trainable params
# Check adapter applied
print(model.peft_config)
```
3. **Check data formatting**:
```python
# Verify tokenization
sample = dataset[0]
decoded = tokenizer.decode(sample["input_ids"])
print(decoded) # Should look correct
```
4. **Increase rank**:
```python
LoraConfig(r=32, lora_alpha=64) # More capacity
```
### NaN Loss
**Error**: `Loss is NaN`
**Fix**:
```python
# Use bf16 instead of fp16
TrainingArguments(bf16=True, fp16=False)
# Or enable loss scaling
TrainingArguments(fp16=True, fp16_full_eval=True)
# Lower learning rate
TrainingArguments(learning_rate=5e-5)
# Check for data issues
for batch in dataloader:
if torch.isnan(batch["input_ids"].float()).any():
print("NaN in input!")
```
### Adapter Not Training
**Problem**: `trainable params: 0` or model not updating.
**Fix**:
```python
# Verify LoRA applied to correct modules
for name, module in model.named_modules():
if "lora" in name.lower():
print(f"Found LoRA: {name}")
# Check target_modules match model architecture
from peft.utils import TRANSFORMERS_MODELS_TO_LORA_TARGET_MODULES_MAPPING
print(TRANSFORMERS_MODELS_TO_LORA_TARGET_MODULES_MAPPING.get(model.config.model_type))
# Ensure model in training mode
model.train()
# Check requires_grad
for name, param in model.named_parameters():
if param.requires_grad:
print(f"Trainable: {name}")
```
## Loading Issues
### Adapter Loading Fails
**Error**: `ValueError: Can't find adapter weights`
**Fix**:
```python
# Check adapter files exist
import os
print(os.listdir("./adapter-path"))
# Should contain: adapter_config.json, adapter_model.safetensors
# Load with correct structure
from peft import PeftModel, PeftConfig
# Check config
config = PeftConfig.from_pretrained("./adapter-path")
print(config)
# Load base model first
base_model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(base_model, "./adapter-path")
```
### Base Model Mismatch
**Error**: `RuntimeError: size mismatch`
**Fix**:
```python
# Ensure base model matches adapter
from peft import PeftConfig
config = PeftConfig.from_pretrained("./adapter-path")
print(f"Base model: {config.base_model_name_or_path}")
# Load exact same base model
base_model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
```
### Safetensors vs PyTorch Format
**Error**: `ValueError: We couldn't connect to 'https://huggingface.co'`
**Fix**:
```python
# Force local loading
model = PeftModel.from_pretrained(
base_model,
"./adapter-path",
local_files_only=True
)
# Or specify format
model.save_pretrained("./adapter", safe_serialization=True) # safetensors
model.save_pretrained("./adapter", safe_serialization=False) # pytorch
```
## Inference Issues
### Slow Generation
**Problem**: Inference much slower than expected.
**Solutions**:
1. **Merge adapter for deployment**:
```python
merged_model = model.merge_and_unload()
# No adapter overhead during inference
```
2. **Use optimized inference engine**:
```python
from vllm import LLM
llm = LLM(model="./merged-model", dtype="half")
```
3. **Enable Flash Attention**:
```python
model = AutoModelForCausalLM.from_pretrained(
model_name,
attn_implementation="flash_attention_2"
)
```
### Output Quality Issues
**Problem**: Fine-tuned model produces worse outputs.
**Solutions**:
1. **Check evaluation without adapter**:
```python
with model.disable_adapter():
base_output = model.generate(**inputs)
# Compare with adapter output
```
2. **Lower temperature during eval**:
```python
model.generate(**inputs, temperature=0.1, do_sample=False)
```
3. **Retrain with more data**:
```python
# Increase training samples
# Use higher quality data
# Train for more epochs
```
### Wrong Adapter Active
**Problem**: Model using wrong adapter or no adapter.
**Fix**:
```python
# Check active adapters
print(model.active_adapters)
# Explicitly set adapter
model.set_adapter("your-adapter-name")
# List all adapters
print(model.peft_config.keys())
```
## QLoRA Specific Issues
### Quantization Errors
**Error**: `RuntimeError: mat1 and mat2 shapes cannot be multiplied`
**Fix**:
```python
# Ensure compute dtype matches
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16, # Match model dtype
bnb_4bit_quant_type="nf4"
)
# Load with correct dtype
model = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=bnb_config,
torch_dtype=torch.bfloat16
)
```
### QLoRA OOM
**Error**: OOM even with 4-bit quantization.
**Fix**:
```python
# Enable double quantization
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True # Further memory reduction
)
# Use offloading
model = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=bnb_config,
device_map="auto",
max_memory={0: "20GB", "cpu": "100GB"}
)
```
### QLoRA Merge Fails
**Error**: `RuntimeError: expected scalar type BFloat16 but found Float`
**Fix**:
```python
# Dequantize before merging
from peft import PeftModel
# Load in higher precision for merging
base_model = AutoModelForCausalLM.from_pretrained(
base_model_name,
torch_dtype=torch.float16, # Not quantized
device_map="auto"
)
# Load adapter
model = PeftModel.from_pretrained(base_model, "./qlora-adapter")
# Now merge
merged = model.merge_and_unload()
```
## Multi-Adapter Issues
### Adapter Conflict
**Error**: `ValueError: Adapter with name 'default' already exists`
**Fix**:
```python
# Use unique names
model.load_adapter("./adapter1", adapter_name="task1")
model.load_adapter("./adapter2", adapter_name="task2")
# Or delete existing
model.delete_adapter("default")
```
### Mixed Precision Adapters
**Error**: Adapters trained with different dtypes.
**Fix**:
```python
# Convert adapter precision
model = PeftModel.from_pretrained(base_model, "./adapter")
model = model.to(torch.bfloat16)
# Or load with specific dtype
model = PeftModel.from_pretrained(
base_model,
"./adapter",
torch_dtype=torch.bfloat16
)
```
## Performance Optimization
### Memory Profiling
```python
import torch
def print_memory():
if torch.cuda.is_available():
allocated = torch.cuda.memory_allocated() / 1e9
reserved = torch.cuda.memory_reserved() / 1e9
print(f"Allocated: {allocated:.2f}GB, Reserved: {reserved:.2f}GB")
# Profile during training
print_memory() # Before
model.train()
loss = model(**batch).loss
loss.backward()
print_memory() # After
```
### Speed Profiling
```python
import time
import torch
def benchmark_generation(model, tokenizer, prompt, n_runs=5):
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
# Warmup
model.generate(**inputs, max_new_tokens=10)
torch.cuda.synchronize()
# Benchmark
times = []
for _ in range(n_runs):
start = time.perf_counter()
outputs = model.generate(**inputs, max_new_tokens=100)
torch.cuda.synchronize()
times.append(time.perf_counter() - start)
tokens = outputs.shape[1] - inputs.input_ids.shape[1]
avg_time = sum(times) / len(times)
print(f"Speed: {tokens/avg_time:.2f} tokens/sec")
# Compare adapter vs merged
benchmark_generation(adapter_model, tokenizer, "Hello")
benchmark_generation(merged_model, tokenizer, "Hello")
```
## Getting Help
1. **Check PEFT GitHub Issues**: https://github.com/huggingface/peft/issues
2. **HuggingFace Forums**: https://discuss.huggingface.co/
3. **PEFT Documentation**: https://huggingface.co/docs/peft
### Debugging Template
When reporting issues, include:
```python
# System info
import peft
import transformers
import torch
print(f"PEFT: {peft.__version__}")
print(f"Transformers: {transformers.__version__}")
print(f"PyTorch: {torch.__version__}")
print(f"CUDA: {torch.version.cuda}")
print(f"GPU: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'N/A'}")
# Config
print(model.peft_config)
model.print_trainable_parameters()
```

129
skills/mlops/training/pytorch-fsdp/SKILL.md Normal file
View File

File diff suppressed because one or more lines are too long

7
skills/mlops/training/pytorch-fsdp/references/index.md Normal file
View File

@@ -0,0 +1,7 @@
# Pytorch-Fsdp Documentation Index
## Categories
### Other
**File:** `other.md`
**Pages:** 15

4261
skills/mlops/training/pytorch-fsdp/references/other.md Normal file
View File

File diff suppressed because it is too large Load Diff

458
skills/mlops/training/trl-fine-tuning/SKILL.md Normal file
View File

@@ -0,0 +1,458 @@
---
name: fine-tuning-with-trl
description: Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.
version: 1.0.0
author: Orchestra Research
license: MIT
dependencies: [trl, transformers, datasets, peft, accelerate, torch]
metadata:
hermes:
tags: [Post-Training, TRL, Reinforcement Learning, Fine-Tuning, SFT, DPO, PPO, GRPO, RLHF, Preference Alignment, HuggingFace]
---
# TRL - Transformer Reinforcement Learning
## Quick start
TRL provides post-training methods for aligning language models with human preferences.
**Installation**:
```bash
pip install trl transformers datasets peft accelerate
```
**Supervised Fine-Tuning** (instruction tuning):
```python
from trl import SFTTrainer
trainer = SFTTrainer(
model="Qwen/Qwen2.5-0.5B",
train_dataset=dataset, # Prompt-completion pairs
)
trainer.train()
```
**DPO** (align with preferences):
```python
from trl import DPOTrainer, DPOConfig
config = DPOConfig(output_dir="model-dpo", beta=0.1)
trainer = DPOTrainer(
model=model,
args=config,
train_dataset=preference_dataset, # chosen/rejected pairs
processing_class=tokenizer
)
trainer.train()
```
## Common workflows
### Workflow 1: Full RLHF pipeline (SFT → Reward Model → PPO)
Complete pipeline from base model to human-aligned model.
Copy this checklist:
```
RLHF Training:
- [ ] Step 1: Supervised fine-tuning (SFT)
- [ ] Step 2: Train reward model
- [ ] Step 3: PPO reinforcement learning
- [ ] Step 4: Evaluate aligned model
```
**Step 1: Supervised fine-tuning**
Train base model on instruction-following data:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
# Load model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")
# Load instruction dataset
dataset = load_dataset("trl-lib/Capybara", split="train")
# Configure training
training_args = SFTConfig(
output_dir="Qwen2.5-0.5B-SFT",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=2e-5,
logging_steps=10,
save_strategy="epoch"
)
# Train
trainer = SFTTrainer(
model=model,
args=training_args,
train_dataset=dataset,
tokenizer=tokenizer
)
trainer.train()
trainer.save_model()
```
**Step 2: Train reward model**
Train model to predict human preferences:
```python
from transformers import AutoModelForSequenceClassification
from trl import RewardTrainer, RewardConfig
# Load SFT model as base
model = AutoModelForSequenceClassification.from_pretrained(
"Qwen2.5-0.5B-SFT",
num_labels=1 # Single reward score
)
tokenizer = AutoTokenizer.from_pretrained("Qwen2.5-0.5B-SFT")
# Load preference data (chosen/rejected pairs)
dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
# Configure training
training_args = RewardConfig(
output_dir="Qwen2.5-0.5B-Reward",
per_device_train_batch_size=2,
num_train_epochs=1,
learning_rate=1e-5
)
# Train reward model
trainer = RewardTrainer(
model=model,
args=training_args,
processing_class=tokenizer,
train_dataset=dataset
)
trainer.train()
trainer.save_model()
```
**Step 3: PPO reinforcement learning**
Optimize policy using reward model:
```bash
python -m trl.scripts.ppo \
--model_name_or_path Qwen2.5-0.5B-SFT \
--reward_model_path Qwen2.5-0.5B-Reward \
--dataset_name trl-internal-testing/descriptiveness-sentiment-trl-style \
--output_dir Qwen2.5-0.5B-PPO \
--learning_rate 3e-6 \
--per_device_train_batch_size 64 \
--total_episodes 10000
```
**Step 4: Evaluate**
```python
from transformers import pipeline
# Load aligned model
generator = pipeline("text-generation", model="Qwen2.5-0.5B-PPO")
# Test
prompt = "Explain quantum computing to a 10-year-old"
output = generator(prompt, max_length=200)[0]["generated_text"]
print(output)
```
### Workflow 2: Simple preference alignment with DPO
Align model with preferences without reward model.
Copy this checklist:
```
DPO Training:
- [ ] Step 1: Prepare preference dataset
- [ ] Step 2: Configure DPO
- [ ] Step 3: Train with DPOTrainer
- [ ] Step 4: Evaluate alignment
```
**Step 1: Prepare preference dataset**
Dataset format:
```json
{
"prompt": "What is the capital of France?",
"chosen": "The capital of France is Paris.",
"rejected": "I don't know."
}
```
Load dataset:
```python
from datasets import load_dataset
dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
# Or load your own
# dataset = load_dataset("json", data_files="preferences.json")
```
**Step 2: Configure DPO**
```python
from trl import DPOConfig
config = DPOConfig(
output_dir="Qwen2.5-0.5B-DPO",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=5e-7,
beta=0.1, # KL penalty strength
max_prompt_length=512,
max_length=1024,
logging_steps=10
)
```
**Step 3: Train with DPOTrainer**
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import DPOTrainer
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
trainer = DPOTrainer(
model=model,
args=config,
train_dataset=dataset,
processing_class=tokenizer
)
trainer.train()
trainer.save_model()
```
**CLI alternative**:
```bash
trl dpo \
--model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
--dataset_name argilla/Capybara-Preferences \
--output_dir Qwen2.5-0.5B-DPO \
--per_device_train_batch_size 4 \
--learning_rate 5e-7 \
--beta 0.1
```
### Workflow 3: Memory-efficient online RL with GRPO
Train with reinforcement learning using minimal memory.
Copy this checklist:
```
GRPO Training:
- [ ] Step 1: Define reward function
- [ ] Step 2: Configure GRPO
- [ ] Step 3: Train with GRPOTrainer
```
**Step 1: Define reward function**
```python
def reward_function(completions, **kwargs):
"""
Compute rewards for completions.
Args:
completions: List of generated texts
Returns:
List of reward scores (floats)
"""
rewards = []
for completion in completions:
# Example: reward based on length and unique words
score = len(completion.split()) # Favor longer responses
score += len(set(completion.lower().split())) # Reward unique words
rewards.append(score)
return rewards
```
Or use a reward model:
```python
from transformers import pipeline
reward_model = pipeline("text-classification", model="reward-model-path")
def reward_from_model(completions, prompts, **kwargs):
# Combine prompt + completion
full_texts = [p + c for p, c in zip(prompts, completions)]
# Get reward scores
results = reward_model(full_texts)
return [r["score"] for r in results]
```
**Step 2: Configure GRPO**
```python
from trl import GRPOConfig
config = GRPOConfig(
output_dir="Qwen2-GRPO",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=1e-5,
num_generations=4, # Generate 4 completions per prompt
max_new_tokens=128
)
```
**Step 3: Train with GRPOTrainer**
```python
from datasets import load_dataset
from trl import GRPOTrainer
# Load prompt-only dataset
dataset = load_dataset("trl-lib/tldr", split="train")
trainer = GRPOTrainer(
model="Qwen/Qwen2-0.5B-Instruct",
reward_funcs=reward_function, # Your reward function
args=config,
train_dataset=dataset
)
trainer.train()
```
**CLI**:
```bash
trl grpo \
--model_name_or_path Qwen/Qwen2-0.5B-Instruct \
--dataset_name trl-lib/tldr \
--output_dir Qwen2-GRPO \
--num_generations 4
```
## When to use vs alternatives
**Use TRL when:**
- Need to align model with human preferences
- Have preference data (chosen/rejected pairs)
- Want to use reinforcement learning (PPO, GRPO)
- Need reward model training
- Doing RLHF (full pipeline)
**Method selection**:
- **SFT**: Have prompt-completion pairs, want basic instruction following
- **DPO**: Have preferences, want simple alignment (no reward model needed)
- **PPO**: Have reward model, need maximum control over RL
- **GRPO**: Memory-constrained, want online RL
- **Reward Model**: Building RLHF pipeline, need to score generations
**Use alternatives instead:**
- **HuggingFace Trainer**: Basic fine-tuning without RL
- **Axolotl**: YAML-based training configuration
- **LitGPT**: Educational, minimal fine-tuning
- **Unsloth**: Fast LoRA training
## Common issues
**Issue: OOM during DPO training**
Reduce batch size and sequence length:
```python
config = DPOConfig(
per_device_train_batch_size=1, # Reduce from 4
max_length=512, # Reduce from 1024
gradient_accumulation_steps=8 # Maintain effective batch
)
```
Or use gradient checkpointing:
```python
model.gradient_checkpointing_enable()
```
**Issue: Poor alignment quality**
Tune beta parameter:
```python
# Higher beta = more conservative (stays closer to reference)
config = DPOConfig(beta=0.5) # Default 0.1
# Lower beta = more aggressive alignment
config = DPOConfig(beta=0.01)
```
**Issue: Reward model not learning**
Check loss type and learning rate:
```python
config = RewardConfig(
learning_rate=1e-5, # Try different LR
num_train_epochs=3 # Train longer
)
```
Ensure preference dataset has clear winners:
```python
# Verify dataset
print(dataset[0])
# Should have clear chosen > rejected
```
**Issue: PPO training unstable**
Adjust KL coefficient:
```python
config = PPOConfig(
kl_coef=0.1, # Increase from 0.05
cliprange=0.1 # Reduce from 0.2
)
```
## Advanced topics
**SFT training guide**: See [references/sft-training.md](references/sft-training.md) for dataset formats, chat templates, packing strategies, and multi-GPU training.
**DPO variants**: See [references/dpo-variants.md](references/dpo-variants.md) for IPO, cDPO, RPO, and other DPO loss functions with recommended hyperparameters.
**Reward modeling**: See [references/reward-modeling.md](references/reward-modeling.md) for outcome vs process rewards, Bradley-Terry loss, and reward model evaluation.
**Online RL methods**: See [references/online-rl.md](references/online-rl.md) for PPO, GRPO, RLOO, and OnlineDPO with detailed configurations.
## Hardware requirements
- **GPU**: NVIDIA (CUDA required)
- **VRAM**: Depends on model and method
- SFT 7B: 16GB (with LoRA)
- DPO 7B: 24GB (stores reference model)
- PPO 7B: 40GB (policy + reward model)
- GRPO 7B: 24GB (more memory efficient)
- **Multi-GPU**: Supported via `accelerate`
- **Mixed precision**: BF16 recommended (A100/H100)
**Memory optimization**:
- Use LoRA/QLoRA for all methods
- Enable gradient checkpointing
- Use smaller batch sizes with gradient accumulation
## Resources
- Docs: https://huggingface.co/docs/trl/
- GitHub: https://github.com/huggingface/trl
- Papers:
- "Training language models to follow instructions with human feedback" (InstructGPT, 2022)
- "Direct Preference Optimization: Your Language Model is Secretly a Reward Model" (DPO, 2023)
- "Group Relative Policy Optimization" (GRPO, 2024)
- Examples: https://github.com/huggingface/trl/tree/main/examples/scripts

227
skills/mlops/training/trl-fine-tuning/references/dpo-variants.md Normal file
View File

@@ -0,0 +1,227 @@
# DPO Variants
Complete guide to Direct Preference Optimization loss variants in TRL.
## Overview
DPO optimizes models using preference data (chosen/rejected pairs). TRL supports 10+ loss variants for different scenarios.
## Loss Types
### 1. Sigmoid (Standard DPO)
**Formula**: `-log(sigmoid(β * logits))`
**When to use**: Default choice, general preference alignment
**Config**:
```python
DPOConfig(
loss_type="sigmoid",
beta=0.1, # KL penalty
per_device_train_batch_size=64,
learning_rate=1e-6
)
```
### 2. IPO (Identity Policy Optimization)
**Formula**: `(logits - 1/(2β))²`
**When to use**: Better theoretical foundation, reduce overfitting
**Config**:
```python
DPOConfig(
loss_type="ipo",
beta=0.1,
per_device_train_batch_size=90,
learning_rate=1e-2
)
```
### 3. Hinge (SLiC)
**Formula**: `ReLU(1 - β * logits)`
**When to use**: Margin-based objective
**Config**:
```python
DPOConfig(
loss_type="hinge",
beta=0.1,
per_device_train_batch_size=512,
learning_rate=1e-4
)
```
### 4. Robust DPO
**Formula**: Sigmoid with label smoothing for noise robustness
**When to use**: Noisy preference labels
**Config**:
```python
DPOConfig(
loss_type="robust",
beta=0.01,
label_smoothing=0.1, # Noise probability
per_device_train_batch_size=16,
learning_rate=1e-3,
max_prompt_length=128,
max_length=512
)
```
### 5. BCO Pair (Binary Classification)
**Formula**: Train binary classifier (chosen=1, rejected=0)
**When to use**: Pairwise preference data
**Config**:
```python
DPOConfig(
loss_type="bco_pair",
beta=0.01,
per_device_train_batch_size=128,
learning_rate=5e-7,
max_prompt_length=1536,
max_completion_length=512
)
```
### 6. SPPO Hard
**Formula**: Push chosen→0.5, rejected→-0.5
**When to use**: Nash equilibrium, sparse data
**Config**:
```python
DPOConfig(
loss_type="sppo_hard",
beta=0.1
)
```
### 7. DiscoPOP
**Formula**: Log-Ratio Modulated Loss
**When to use**: Automated loss discovery
**Config**:
```python
DPOConfig(
loss_type="discopop",
beta=0.05,
discopop_tau=0.05,
per_device_train_batch_size=64,
learning_rate=5e-7
)
```
### 8. APO Zero
**Formula**: Increase chosen, decrease rejected likelihood
**When to use**: Model worse than winning outputs
**Config**:
```python
DPOConfig(
loss_type="apo_zero",
beta=0.1,
per_device_train_batch_size=64,
learning_rate=2e-7,
max_prompt_length=512,
max_completion_length=512
)
```
### 9. APO Down
**Formula**: Decrease both, emphasize rejected reduction
**When to use**: Model better than winning outputs
**Config**:
```python
DPOConfig(
loss_type="apo_down",
beta=0.1,
# Same hyperparameters as apo_zero
)
```
### 10. AOT & AOT Pair
**Formula**: Distributional alignment via stochastic dominance
**When to use**:
- `aot_pair`: Paired preference data
- `aot`: Unpaired data
**Config**:
```python
DPOConfig(
loss_type="aot_pair", # or "aot"
beta=0.1,
label_smoothing=0.0
)
```
## Multi-Loss Training
Combine multiple losses:
```python
DPOConfig(
loss_type=["sigmoid", "ipo"],
loss_weights=[0.7, 0.3], # Weighted combination
beta=0.1
)
```
## Key Parameters
### Beta (β)
Controls deviation from reference model:
- **Higher** (0.5): More conservative, stays close to reference
- **Lower** (0.01): More aggressive alignment
- **Default**: 0.1
### Label Smoothing
For robust DPO:
- **0.0**: No smoothing (default)
- **0.1-0.3**: Moderate noise robustness
- **0.5**: Maximum noise tolerance
### Max Lengths
- `max_prompt_length`: 128-1536
- `max_completion_length`: 128-512
- `max_length`: Total sequence (1024-2048)
## Comparison Table
| Loss | Speed | Stability | Best For |
|------|-------|-----------|----------|
| Sigmoid | Fast | Good | **General use** |
| IPO | Fast | Better | Overfitting issues |
| Hinge | Fast | Good | Margin objectives |
| Robust | Fast | Best | Noisy data |
| BCO | Medium | Good | Binary classification |
| DiscoPOP | Fast | Good | New architectures |
| APO | Fast | Good | Model quality matching |
## References
- DPO paper: https://arxiv.org/abs/2305.18290
- IPO paper: https://arxiv.org/abs/2310.12036
- TRL docs: https://huggingface.co/docs/trl/dpo_trainer

82
skills/mlops/training/trl-fine-tuning/references/online-rl.md Normal file
View File

@@ -0,0 +1,82 @@
# Online RL Methods
Guide to online reinforcement learning with PPO, GRPO, RLOO, and OnlineDPO.
## Overview
Online RL generates completions during training and optimizes based on rewards.
## PPO (Proximal Policy Optimization)
Classic RL algorithm for LLM alignment.
### Basic Usage
```bash
python -m trl.scripts.ppo \
--model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
--reward_model_path reward-model \
--dataset_name trl-internal-testing/descriptiveness-sentiment-trl-style \
--output_dir model-ppo \
--learning_rate 3e-6 \
--per_device_train_batch_size 64 \
--total_episodes 10000 \
--num_ppo_epochs 4 \
--kl_coef 0.05
```
### Key Parameters
- `kl_coef`: KL penalty (0.05-0.2)
- `num_ppo_epochs`: Epochs per batch (2-4)
- `cliprange`: PPO clip (0.1-0.3)
- `vf_coef`: Value function coef (0.1)
## GRPO (Group Relative Policy Optimization)
Memory-efficient online RL.
### Basic Usage
```python
from trl import GRPOTrainer, GRPOConfig
from datasets import load_dataset
# Define reward function
def reward_func(completions, **kwargs):
return [len(set(c.split())) for c in completions]
config = GRPOConfig(
output_dir="model-grpo",
num_generations=4, # Completions per prompt
max_new_tokens=128
)
trainer = GRPOTrainer(
model="Qwen/Qwen2-0.5B-Instruct",
reward_funcs=reward_func,
args=config,
train_dataset=load_dataset("trl-lib/tldr", split="train")
)
trainer.train()
```
### Key Parameters
- `num_generations`: 2-8 completions
- `max_new_tokens`: 64-256
- Learning rate: 1e-5 to 1e-4
## Memory Comparison
| Method | Memory (7B) | Speed | Use Case |
|--------|-------------|-------|----------|
| PPO | 40GB | Medium | Maximum control |
| GRPO | 24GB | Fast | **Memory-constrained** |
| OnlineDPO | 28GB | Fast | No reward model |
## References
- PPO paper: https://arxiv.org/abs/1707.06347
- GRPO paper: https://arxiv.org/abs/2402.03300
- TRL docs: https://huggingface.co/docs/trl/

122
skills/mlops/training/trl-fine-tuning/references/reward-modeling.md Normal file
View File

@@ -0,0 +1,122 @@
# Reward Modeling
Guide to training reward models with TRL for RLHF pipelines.
## Overview
Reward models score completions based on human preferences. Used in:
- PPO training (RL feedback)
- GRPO online RL
- Completion ranking
## Basic Training
```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from trl import RewardTrainer, RewardConfig
from datasets import load_dataset
# Load model (num_labels=1 for single reward score)
model = AutoModelForSequenceClassification.from_pretrained(
"Qwen/Qwen2.5-0.5B-Instruct",
num_labels=1
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
# Load preference dataset (chosen/rejected pairs)
dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
# Configure
config = RewardConfig(
output_dir="Qwen2.5-Reward",
per_device_train_batch_size=2,
num_train_epochs=1,
learning_rate=1e-5
)
# Train
trainer = RewardTrainer(
model=model,
args=config,
processing_class=tokenizer,
train_dataset=dataset
)
trainer.train()
```
## Dataset Format
Required fields:
```json
{
"prompt": "Question or instruction",
"chosen": "Better response",
"rejected": "Worse response"
}
```
## Bradley-Terry Loss
Default loss function:
```
loss = -log(sigmoid(reward_chosen - reward_rejected))
```
Learns to score chosen > rejected.
## Using Reward Models
### Inference
```python
from transformers import pipeline
# Load trained reward model
reward_pipe = pipeline("text-classification", model="Qwen2.5-Reward")
# Score completions
texts = ["Good answer", "Bad answer"]
scores = reward_pipe(texts)
print(scores) # Higher score = better
```
### In PPO
```python
from trl import PPOTrainer, PPOConfig
config = PPOConfig(
reward_model_path="Qwen2.5-Reward" # Use trained reward model
)
trainer = PPOTrainer(
model=policy_model,
config=config,
# Reward model loaded automatically
)
```
## Hyperparameters
| Model Size | Learning Rate | Batch Size | Epochs |
|------------|---------------|------------|--------|
| <1B | 2e-5 | 4-8 | 1-2 |
| 1-7B | 1e-5 | 2-4 | 1 |
| 7-13B | 5e-6 | 1-2 | 1 |
## Evaluation
Check reward separation:
```python
# Chosen should score higher than rejected
chosen_rewards = model(**chosen_inputs).logits
rejected_rewards = model(**rejected_inputs).logits
accuracy = (chosen_rewards > rejected_rewards).float().mean()
print(f"Accuracy: {accuracy:.2%}") # Target: >80%
```
## References
- InstructGPT paper: https://arxiv.org/abs/2203.02155
- TRL docs: https://huggingface.co/docs/trl/reward_trainer

168
skills/mlops/training/trl-fine-tuning/references/sft-training.md Normal file
View File

@@ -0,0 +1,168 @@
# SFT Training Guide
Complete guide to Supervised Fine-Tuning (SFT) with TRL for instruction tuning and task-specific fine-tuning.
## Overview
SFT trains models on input-output pairs to minimize cross-entropy loss. Use for:
- Instruction following
- Task-specific fine-tuning
- Chatbot training
- Domain adaptation
## Dataset Formats
### Format 1: Prompt-Completion
```json
[
{
"prompt": "What is the capital of France?",
"completion": "The capital of France is Paris."
}
]
```
### Format 2: Conversational (ChatML)
```json
[
{
"messages": [
{"role": "user", "content": "What is Python?"},
{"role": "assistant", "content": "Python is a programming language."}
]
}
]
```
### Format 3: Text-only
```json
[
{"text": "User: Hello\nAssistant: Hi! How can I help?"}
]
```
## Basic Training
```python
from trl import SFTTrainer, SFTConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset
# Load model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")
# Load dataset
dataset = load_dataset("trl-lib/Capybara", split="train")
# Configure
config = SFTConfig(
output_dir="Qwen2.5-SFT",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=2e-5,
save_strategy="epoch"
)
# Train
trainer = SFTTrainer(
model=model,
args=config,
train_dataset=dataset,
tokenizer=tokenizer
)
trainer.train()
```
## Chat Templates
Apply chat templates automatically:
```python
trainer = SFTTrainer(
model=model,
args=config,
train_dataset=dataset, # Messages format
tokenizer=tokenizer
# Chat template applied automatically
)
```
Or manually:
```python
def format_chat(example):
messages = example["messages"]
text = tokenizer.apply_chat_template(messages, tokenize=False)
return {"text": text}
dataset = dataset.map(format_chat)
```
## Packing for Efficiency
Pack multiple sequences into one to maximize GPU utilization:
```python
config = SFTConfig(
packing=True, # Enable packing
max_seq_length=2048,
dataset_text_field="text"
)
```
**Benefits**: 2-3× faster training
**Trade-off**: Slightly more complex batching
## Multi-GPU Training
```bash
accelerate launch --num_processes 4 train_sft.py
```
Or with config:
```python
config = SFTConfig(
output_dir="model-sft",
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
num_train_epochs=1
)
```
## LoRA Fine-Tuning
```python
from peft import LoraConfig
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules="all-linear",
lora_dropout=0.05,
task_type="CAUSAL_LM"
)
trainer = SFTTrainer(
model=model,
args=config,
train_dataset=dataset,
peft_config=lora_config # Add LoRA
)
```
## Hyperparameters
| Model Size | Learning Rate | Batch Size | Epochs |
|------------|---------------|------------|--------|
| <1B | 5e-5 | 8-16 | 1-3 |
| 1-7B | 2e-5 | 4-8 | 1-2 |
| 7-13B | 1e-5 | 2-4 | 1 |
| 13B+ | 5e-6 | 1-2 | 1 |
## References
- TRL docs: https://huggingface.co/docs/trl/sft_trainer
- Examples: https://github.com/huggingface/trl/tree/main/examples/scripts

83
skills/mlops/training/unsloth/SKILL.md Normal file
View File

@@ -0,0 +1,83 @@
---
name: unsloth
description: Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA/QLoRA optimization
version: 1.0.0
author: Orchestra Research
license: MIT
dependencies: [unsloth, torch, transformers, trl, datasets, peft]
metadata:
hermes:
tags: [Fine-Tuning, Unsloth, Fast Training, LoRA, QLoRA, Memory-Efficient, Optimization, Llama, Mistral, Gemma, Qwen]
---
# Unsloth Skill
Comprehensive assistance with unsloth development, generated from official documentation.
## When to Use This Skill
This skill should be triggered when:
- Working with unsloth
- Asking about unsloth features or APIs
- Implementing unsloth solutions
- Debugging unsloth code
- Learning unsloth best practices
## Quick Reference
### Common Patterns
*Quick reference patterns will be added as you use the skill.*
## Reference Files
This skill includes comprehensive documentation in `references/`:
- **llms-txt.md** - Llms-Txt documentation
Use `view` to read specific reference files when detailed information is needed.
## Working with This Skill
### For Beginners
Start with the getting_started or tutorials reference files for foundational concepts.
### For Specific Features
Use the appropriate category reference file (api, guides, etc.) for detailed information.
### For Code Examples
The quick reference section above contains common patterns extracted from the official docs.
## Resources
### references/
Organized documentation extracted from official sources. These files contain:
- Detailed explanations
- Code examples with language annotations
- Links to original documentation
- Table of contents for quick navigation
### scripts/
Add helper scripts here for common automation tasks.
### assets/
Add templates, boilerplate, or example projects here.
## Notes
- This skill was automatically generated from official documentation
- Reference files preserve the structure and examples from source docs
- Code examples include language detection for better syntax highlighting
- Quick reference patterns are extracted from common usage examples in the docs
## Updating
To refresh this skill with updated documentation:
1. Re-run the scraper with the same configuration
2. The skill will be rebuilt with the latest information
<!-- Trigger re-upload 1763621536 -->

7
skills/mlops/training/unsloth/references/index.md Normal file
View File

@@ -0,0 +1,7 @@
# Unsloth Documentation Index
## Categories
### Llms-Txt
**File:** `llms-txt.md`
**Pages:** 136

16799
skills/mlops/training/unsloth/references/llms-full.md Normal file
View File

File diff suppressed because one or more lines are too long

12044
skills/mlops/training/unsloth/references/llms-txt.md Normal file
View File

File diff suppressed because one or more lines are too long

82
skills/mlops/training/unsloth/references/llms.md Normal file
View File

@@ -0,0 +1,82 @@
# Unsloth Documentation
## Unsloth Documentation
- [Unsloth Docs](/get-started/unsloth-docs.md): Train your own model with Unsloth, an open-source framework for LLM fine-tuning and reinforcement learning.
- [Beginner? Start here!](/get-started/beginner-start-here.md)
- [Unsloth Requirements](/get-started/beginner-start-here/unsloth-requirements.md): Here are Unsloth's requirements including system and GPU VRAM requirements.
- [FAQ + Is Fine-tuning Right For Me?](/get-started/beginner-start-here/faq-+-is-fine-tuning-right-for-me.md): If you're stuck on if fine-tuning is right for you, see here! Learn about fine-tuning misconceptions, how it compared to RAG and more:
- [Unsloth Notebooks](/get-started/unsloth-notebooks.md): Explore our catalog of Unsloth notebooks:
- [All Our Models](/get-started/all-our-models.md)
- [Install & Update](/get-started/install-and-update.md): Learn to install Unsloth locally or online.
- [Updating](/get-started/install-and-update/updating.md): To update or use an old version of Unsloth, follow the steps below:
- [Pip Install](/get-started/install-and-update/pip-install.md): To install Unsloth locally via Pip, follow the steps below:
- [Docker](/get-started/install-and-update/docker.md): Install Unsloth using our official Docker container
- [Windows Installation](/get-started/install-and-update/windows-installation.md): See how to install Unsloth on Windows with or without WSL.
- [AMD](/get-started/install-and-update/amd.md): Fine-tune with Unsloth on AMD GPUs.
- [Conda Install](/get-started/install-and-update/conda-install.md): To install Unsloth locally on Conda, follow the steps below:
- [Google Colab](/get-started/install-and-update/google-colab.md): To install and run Unsloth on Google Colab, follow the steps below:
- [Fine-tuning LLMs Guide](/get-started/fine-tuning-llms-guide.md): Learn all the basics and best practices of fine-tuning. Beginner-friendly.
- [What Model Should I Use?](/get-started/fine-tuning-llms-guide/what-model-should-i-use.md)
- [Datasets Guide](/get-started/fine-tuning-llms-guide/datasets-guide.md): Learn how to create & prepare a dataset for fine-tuning.
- [LoRA Hyperparameters Guide](/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide.md): Optimal lora rank. alpha, number of epochs, batch size & gradient accumulation, QLoRA vs LoRA, target modules and more!
- [Tutorial: How to Finetune Llama-3 and Use In Ollama](/get-started/fine-tuning-llms-guide/tutorial-how-to-finetune-llama-3-and-use-in-ollama.md): Beginner's Guide for creating a customized personal assistant (like ChatGPT) to run locally on Ollama
- [Reinforcement Learning (RL) Guide](/get-started/reinforcement-learning-rl-guide.md): Learn all about Reinforcement Learning (RL) and how to train your own DeepSeek-R1 reasoning model with Unsloth using GRPO. A complete guide from beginner to advanced.
- [Tutorial: Train your own Reasoning model with GRPO](/get-started/reinforcement-learning-rl-guide/tutorial-train-your-own-reasoning-model-with-grpo.md): Beginner's Guide to transforming a model like Llama 3.1 (8B) into a reasoning model by using Unsloth and GRPO.
- [Advanced RL Documentation](/get-started/reinforcement-learning-rl-guide/advanced-rl-documentation.md): Advanced documentation settings when using Unsloth with GRPO.
- [Memory Efficient RL](/get-started/reinforcement-learning-rl-guide/memory-efficient-rl.md)
- [RL Reward Hacking](/get-started/reinforcement-learning-rl-guide/rl-reward-hacking.md): Learn what is Reward Hacking in Reinforcement Learning and how to counter it.
- [GSPO Reinforcement Learning](/get-started/reinforcement-learning-rl-guide/gspo-reinforcement-learning.md): Train with GSPO (Group Sequence Policy Optimization) RL in Unsloth.
- [Reinforcement Learning - DPO, ORPO & KTO](/get-started/reinforcement-learning-rl-guide/reinforcement-learning-dpo-orpo-and-kto.md): To use the reward modelling functions for DPO, GRPO, ORPO or KTO with Unsloth, follow the steps below:
- [DeepSeek-OCR: How to Run & Fine-tune](/new/deepseek-ocr-how-to-run-and-fine-tune.md): Guide on how to run and fine-tune DeepSeek-OCR locally.
- [How to Fine-tune LLMs with Unsloth & Docker](/new/how-to-fine-tune-llms-with-unsloth-and-docker.md): Learn how to fine-tune LLMs or do Reinforcement Learning (RL) with Unsloth's Docker image.
- [Vision Reinforcement Learning (VLM RL)](/new/vision-reinforcement-learning-vlm-rl.md): Train Vision/multimodal models via GRPO and RL with Unsloth!
- [gpt-oss Reinforcement Learning](/new/gpt-oss-reinforcement-learning.md)
- [Tutorial: How to Train gpt-oss with RL](/new/gpt-oss-reinforcement-learning/tutorial-how-to-train-gpt-oss-with-rl.md): Learn to train OpenAI gpt-oss with GRPO to autonomously beat 2048 locally or on Colab.
- [Unsloth Dynamic GGUFs on Aider Polyglot](/new/unsloth-dynamic-ggufs-on-aider-polyglot.md): Performance of Unsloth Dynamic GGUFs on Aider Polyglot Benchmarks
- [Qwen3-VL: How to Run & Fine-tune](/models/qwen3-vl-how-to-run-and-fine-tune.md): Learn to fine-tune and run Qwen3-VL locally with Unsloth.
- [gpt-oss: How to Run & Fine-tune](/models/gpt-oss-how-to-run-and-fine-tune.md): Run & fine-tune OpenAI's new open-source models!
- [Tutorial: How to Fine-tune gpt-oss](/models/gpt-oss-how-to-run-and-fine-tune/tutorial-how-to-fine-tune-gpt-oss.md): Learn step-by-step how to train OpenAI gpt-oss locally with Unsloth.
- [Long Context gpt-oss Training](/models/gpt-oss-how-to-run-and-fine-tune/long-context-gpt-oss-training.md)
- [GLM-4.6: How to Run Locally](/models/glm-4.6-how-to-run-locally.md): A guide on how to run Z.ai's new GLM-4.6 model on your own local device!
- [IBM Granite 4.0](/models/ibm-granite-4.0.md): How to run IBM Granite-4.0 with Unsloth GGUFs on llama.cpp, Ollama and how to fine-tune!
- [DeepSeek-V3.1: How to Run Locally](/models/deepseek-v3.1-how-to-run-locally.md): A guide on how to run DeepSeek-V3.1 and Terminus on your own local device!
- [Qwen3-Coder: How to Run Locally](/models/qwen3-coder-how-to-run-locally.md): Run Qwen3-Coder-30B-A3B-Instruct and 480B-A35B locally with Unsloth Dynamic quants.
- [Gemma 3: How to Run & Fine-tune](/models/gemma-3-how-to-run-and-fine-tune.md): How to run Gemma 3 effectively with our GGUFs on llama.cpp, Ollama, Open WebUI and how to fine-tune with Unsloth!
- [Gemma 3n: How to Run & Fine-tune](/models/gemma-3-how-to-run-and-fine-tune/gemma-3n-how-to-run-and-fine-tune.md): Run Google's new Gemma 3n locally with Dynamic GGUFs on llama.cpp, Ollama, Open WebUI and fine-tune with Unsloth!
- [Qwen3: How to Run & Fine-tune](/models/qwen3-how-to-run-and-fine-tune.md): Learn to run & fine-tune Qwen3 locally with Unsloth + our Dynamic 2.0 quants
- [Qwen3-2507](/models/qwen3-how-to-run-and-fine-tune/qwen3-2507.md): Run Qwen3-30B-A3B-2507 and 235B-A22B Thinking and Instruct versions locally on your device!
- [Tutorials: How To Fine-tune & Run LLMs](/models/tutorials-how-to-fine-tune-and-run-llms.md): Learn how to run and fine-tune models for optimal performance 100% locally with Unsloth.
- [DeepSeek-R1-0528: How to Run Locally](/models/tutorials-how-to-fine-tune-and-run-llms/deepseek-r1-0528-how-to-run-locally.md): A guide on how to run DeepSeek-R1-0528 including Qwen3 on your own local device!
- [Magistral: How to Run & Fine-tune](/models/tutorials-how-to-fine-tune-and-run-llms/magistral-how-to-run-and-fine-tune.md): Meet Magistral - Mistral's new reasoning models.
- [Llama 4: How to Run & Fine-tune](/models/tutorials-how-to-fine-tune-and-run-llms/llama-4-how-to-run-and-fine-tune.md): How to run Llama 4 locally using our dynamic GGUFs which recovers accuracy compared to standard quantization.
- [Kimi K2: How to Run Locally](/models/tutorials-how-to-fine-tune-and-run-llms/kimi-k2-how-to-run-locally.md): Guide on running Kimi K2 and Kimi-K2-Instruct-0905 on your own local device!
- [Grok 2](/models/tutorials-how-to-fine-tune-and-run-llms/grok-2.md): Run xAI's Grok 2 model locally!
- [Devstral: How to Run & Fine-tune](/models/tutorials-how-to-fine-tune-and-run-llms/devstral-how-to-run-and-fine-tune.md): Run and fine-tune Mistral Devstral 1.1, including Small-2507 and 2505.
- [DeepSeek-V3-0324: How to Run Locally](/models/tutorials-how-to-fine-tune-and-run-llms/deepseek-v3-0324-how-to-run-locally.md): How to run DeepSeek-V3-0324 locally using our dynamic quants which recovers accuracy
- [DeepSeek-R1: How to Run Locally](/models/tutorials-how-to-fine-tune-and-run-llms/deepseek-r1-how-to-run-locally.md): A guide on how you can run our 1.58-bit Dynamic Quants for DeepSeek-R1 using llama.cpp.
- [DeepSeek-R1 Dynamic 1.58-bit](/models/tutorials-how-to-fine-tune-and-run-llms/deepseek-r1-how-to-run-locally/deepseek-r1-dynamic-1.58-bit.md): See performance comparison tables for Unsloth's Dynamic GGUF Quants vs Standard IMatrix Quants.
- [QwQ-32B: How to Run effectively](/models/tutorials-how-to-fine-tune-and-run-llms/qwq-32b-how-to-run-effectively.md): How to run QwQ-32B effectively with our bug fixes and without endless generations + GGUFs.
- [Phi-4 Reasoning: How to Run & Fine-tune](/models/tutorials-how-to-fine-tune-and-run-llms/phi-4-reasoning-how-to-run-and-fine-tune.md): Learn to run & fine-tune Phi-4 reasoning models locally with Unsloth + our Dynamic 2.0 quants
- [Running & Saving Models](/basics/running-and-saving-models.md): Learn how to save your finetuned model so you can run it in your favorite inference engine.
- [Saving to GGUF](/basics/running-and-saving-models/saving-to-gguf.md): Saving models to 16bit for GGUF so you can use it for Ollama, Jan AI, Open WebUI and more!
- [Saving to Ollama](/basics/running-and-saving-models/saving-to-ollama.md)
- [Saving to vLLM for deployment](/basics/running-and-saving-models/saving-to-vllm-for-deployment.md): Saving models to 16bit for vLLM deployment and serving
- [Saving to SGLang for deployment](/basics/running-and-saving-models/saving-to-sglang-for-deployment.md): Saving models to 16bit for SGLang for deployment and serving
- [Unsloth Inference](/basics/running-and-saving-models/unsloth-inference.md): Learn how to run your finetuned model with Unsloth's faster inference.
- [Troubleshooting Inference](/basics/running-and-saving-models/troubleshooting-inference.md): If you're experiencing issues when running or saving your model.
- [vLLM Engine Arguments](/basics/running-and-saving-models/vllm-engine-arguments.md)
- [LoRA Hot Swapping Guide](/basics/running-and-saving-models/lora-hot-swapping-guide.md)
- [Text-to-Speech (TTS) Fine-tuning](/basics/text-to-speech-tts-fine-tuning.md): Learn how to fine-tune TTS & STT voice models with Unsloth.
- [Unsloth Dynamic 2.0 GGUFs](/basics/unsloth-dynamic-2.0-ggufs.md): A big new upgrade to our Dynamic Quants!
- [Vision Fine-tuning](/basics/vision-fine-tuning.md): Learn how to fine-tune vision/multimodal LLMs with Unsloth
- [Fine-tuning LLMs with NVIDIA DGX Spark and Unsloth](/basics/fine-tuning-llms-with-nvidia-dgx-spark-and-unsloth.md): Tutorial on how to fine-tune and do reinforcement learning (RL) with OpenAI gpt-oss on NVIDIA DGX Spark.
- [Fine-tuning LLMs with Blackwell, RTX 50 series & Unsloth](/basics/fine-tuning-llms-with-blackwell-rtx-50-series-and-unsloth.md): Learn how to fine-tune LLMs on NVIDIA's Blackwell RTX 50 series and B200 GPUs with our step-by-step guide.
- [Multi-GPU Training with Unsloth](/basics/multi-gpu-training-with-unsloth.md): Learn how to fine-tune LLMs on multiple GPUs and parallelism with Unsloth.
- [Finetuning from Last Checkpoint](/basics/finetuning-from-last-checkpoint.md): Checkpointing allows you to save your finetuning progress so you can pause it and then continue.
- [Troubleshooting & FAQs](/basics/troubleshooting-and-faqs.md): Tips to solve issues, and frequently asked questions.
- [Chat Templates](/basics/chat-templates.md): Learn the fundamentals and customization options of chat templates, including Conversational, ChatML, ShareGPT, Alpaca formats, and more!
- [Quantization-Aware Training (QAT)](/basics/quantization-aware-training-qat.md): Quantize models to 4-bit with Unsloth and PyTorch to recover accuracy.
- [Unsloth Environment Flags](/basics/unsloth-environment-flags.md): Advanced flags which might be useful if you see breaking finetunes, or you want to turn stuff off.
- [Continued Pretraining](/basics/continued-pretraining.md): AKA as Continued Finetuning. Unsloth allows you to continually pretrain so a model can learn a new language.
- [Unsloth Benchmarks](/basics/unsloth-benchmarks.md): Unsloth recorded benchmarks on NVIDIA GPUs.