metadata

language:
  - en
license: other
library_name: transformers
tags:
  - orpo
  - llama 3
  - rlhf
  - sft
datasets:
  - mlabonne/orpo-dpo-mix-40k

OrpoLlama-3-8B

This is an ORPO fine-tune of meta-llama/Meta-Llama-3-8B on 1k samples of mlabonne/orpo-dpo-mix-40k created for this article.

It's a successful fine-tune that follows the ChatML template!

Try the demo: https://huggingface.co/spaces/mlabonne/OrpoLlama-3-8B

🔎 Application

This model uses a context window of 8k. It was trained with the ChatML template.

🏆 Evaluation

Nous

OrpoLlama-4-8B outperforms Llama-3-8B-Instruct on the GPT4All and TruthfulQA datasets.

Evaluation performed using LLM AutoEval, see the entire leaderboard here.

Model	Average	AGIEval	GPT4All	TruthfulQA	Bigbench
meta-llama/Meta-Llama-3-8B-Instruct 📄	51.34	41.22	69.86	51.65	42.64
mlabonne/OrpoLlama-3-8B 📄	48.63	34.17	70.59	52.39	37.36
mlabonne/OrpoLlama-3-8B-1k 📄	46.76	31.56	70.19	48.11	37.17
meta-llama/Meta-Llama-3-8B 📄	45.42	31.1	69.95	43.91	36.7

mlabonne/OrpoLlama-3-8B-1k corresponds to a version of this model trained on 1K samples (you can see the parameters in this article).

Open LLM Leaderboard

TBD.

📈 Training curves

You can find the experiment on W&B at this address.

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/OrpoLlama-3-8B"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])