|
|
--- |
|
|
base_model: meta-llama/Llama-3.2-1B-Instruct |
|
|
tags: |
|
|
- dpo |
|
|
- preference-learning |
|
|
- llm-judge |
|
|
- peft |
|
|
- lora |
|
|
license: llama3.2 |
|
|
language: |
|
|
- en |
|
|
--- |
|
|
|
|
|
# Llama-3.2-1B DPO Fine-tuned (LLM Judge) |
|
|
|
|
|
This model is a DPO (Direct Preference Optimization) fine-tuned version of Llama-3.2-1B-Instruct, |
|
|
trained on preference data generated using an LLM judge system. |
|
|
|
|
|
## Training Details |
|
|
|
|
|
- **Base Model**: meta-llama/Llama-3.2-1B-Instruct |
|
|
- **Training Method**: DPO (Direct Preference Optimization) |
|
|
- **Dataset**: LLM Judge preference pairs (15 samples) |
|
|
- **LoRA Configuration**: r=16, alpha=32 |
|
|
- **Training Epochs**: 3 |
|
|
- **Beta (DPO temperature)**: 0.1 |
|
|
- **Learning Rate**: 5e-5 |
|
|
|
|
|
## Preference Collection Method |
|
|
|
|
|
The training dataset was created using an LLM-based judge system that evaluates responses based on: |
|
|
- Helpfulness |
|
|
- Accuracy |
|
|
- Safety |
|
|
- Coherence |
|
|
- Conciseness |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
from peft import PeftModel |
|
|
|
|
|
# Load base model |
|
|
base_model = AutoModelForCausalLM.from_pretrained( |
|
|
"meta-llama/Llama-3.2-1B-Instruct", |
|
|
device_map="auto" |
|
|
) |
|
|
|
|
|
# Load LoRA adapter |
|
|
model = PeftModel.from_pretrained(base_model, "Zickl/llama32-1b-dpo-llm-judge") |
|
|
|
|
|
# Load tokenizer |
|
|
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B-Instruct") |
|
|
|
|
|
# Generate |
|
|
messages = [{"role": "user", "content": "Your question here"}] |
|
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
|
|
outputs = model.generate(**inputs, max_new_tokens=256) |
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
## Training Logs |
|
|
|
|
|
- Agreement Rate (LLM Judge vs PairRM): 93.3% |
|
|
- Training completed successfully with stable loss convergence |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Trained on small dataset (15 preference pairs) |
|
|
- May exhibit judge biases |
|
|
- Optimized for specific evaluation criteria |
|
|
- 1B parameter model has inherent capability limits |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use this model, please cite: |
|
|
``` |
|
|
@misc{llama32-dpo-llm-judge, |
|
|
author = {Zickl}, |
|
|
title = {Llama-3.2-1B DPO Fine-tuned with LLM Judge}, |
|
|
year = {2024}, |
|
|
publisher = {HuggingFace}, |
|
|
url = {https://huggingface.co/Zickl/llama32-1b-dpo-llm-judge} |
|
|
} |
|
|
``` |
|
|
|
