Mistral-ORPO is a fine-tuned version of mistralai/Mistral-7B-v0.1 using the odds ratio preference optimization (ORPO). With ORPO, the model directly learns the preference without the supervised fine-tuning warmup phase. Mistral-ORPO-⍺ is fine-tuned exclusively on HuggingFaceH4/ultrafeedback_binarized.

Github Repository: https://github.com/xfactlab/orpo

👍 Model Performance

1) AlpacaEval & MT-Bench

Model Name	Size	Align	MT-Bench	AlpacaEval 1.0	AlpacaEval 2.0
Mistral-`ORPO`-⍺	7B	`ORPO`	7.23	87.92	11.33
Mistral-`ORPO`-β	7B	`ORPO`	7.32	91.41	12.20
Zephyr β	7B	DPO	7.34	90.60	10.99
TULU-2-DPO	13B	DPO	7.00	89.5	10.12
Llama-2-Chat	7B	RLHF	6.27	71.37	4.96
Llama-2-Chat	13B	RLHF	6.65	81.09	7.70

2) IFEval

Model Type	Prompt-Strict	Prompt-Loose	Inst-Strict	Inst-Loose
Mistral-ORPO-⍺	0.5009	0.5083	0.5995	0.6163
Mistral-ORPO-β	0.5287	0.5564	0.6355	0.6619

🗺️ MT-Bench by Category

🖥️ Inference

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("kaist-ai/mistral-orpo-alpha")
tokenizer = AutoTokenizer.from_pretrained("kaist-ai/mistral-orpo-alpha")

# Apply chat template
query = [{'role': 'user', 'content': 'Hi! How are you doing?'}]
prompt = tokenizer.apply_chat_template(query, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors='pt')

# Generation with specific configurations
output = model.generate(
  **inputs,
  max_new_tokens=128,
  do_sample=True,
  temperature=0.7
)
response = tokenizer.batch_decode(output)

#<|user|>
#Hi! How are you doing?</s>
#<|assistant|>
#I'm doing well, thank you! How are you?</s>

📎 Citation

@misc{hong2024orpo,
      title={ORPO: Monolithic Preference Optimization without Reference Model}, 
      author={Jiwoo Hong and Noah Lee and James Thorne},
      year={2024},
      eprint={2403.07691},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}