README.md

metadata

language:
  - en
license: llama3
library_name: transformers
tags:
  - orpo
  - llama 3
  - rlhf
  - sft
base_model:
  - meta-llama/Meta-Llama-3-70B
datasets:
  - mlabonne/orpo-dpo-mix-40k
model-index:
  - name: Llama-3-70B-Orpo-v0.1
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: IFEval (0-Shot)
          type: HuggingFaceH4/ifeval
          args:
            num_few_shot: 0
        metrics:
          - type: inst_level_strict_acc and prompt_level_strict_acc
            value: 20.49
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=dfurman/Llama-3-70B-Orpo-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: BBH (3-Shot)
          type: BBH
          args:
            num_few_shot: 3
        metrics:
          - type: acc_norm
            value: 24.09
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=dfurman/Llama-3-70B-Orpo-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MATH Lvl 5 (4-Shot)
          type: hendrycks/competition_math
          args:
            num_few_shot: 4
        metrics:
          - type: exact_match
            value: 13.52
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=dfurman/Llama-3-70B-Orpo-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GPQA (0-shot)
          type: Idavidrein/gpqa
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 1.01
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=dfurman/Llama-3-70B-Orpo-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MuSR (0-shot)
          type: TAUR-Lab/MuSR
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 16.28
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=dfurman/Llama-3-70B-Orpo-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU-PRO (5-shot)
          type: TIGER-Lab/MMLU-Pro
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 32.14
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=dfurman/Llama-3-70B-Orpo-v0.1
          name: Open LLM Leaderboard

dfurman/Llama-3-70B-Orpo-v0.1

This is an ORPO fine-tune of meta-llama/Meta-Llama-3-70B on 2k samples of mlabonne/orpo-dpo-mix-40k.

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

🔎 Application

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

🏆 Evaluation

Open LLM Leaderboard

Model ID	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
meta-llama/Meta-Llama-3-70B-Instruct 📄	77.88	71.42	85.69	80.06	61.81	82.87	85.44
dfurman/Llama-3-70B-Orpo-v0.1 📄	74.67	68.69	88.01	79.39	49.62	85.48	76.8
meta-llama/Meta-Llama-3-70B 📄	73.96	68.77	87.98	79.23	45.56	85.32	76.88

📈 Training curves

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

💻 Usage

Setup

!pip install -qU transformers accelerate bitsandbytes

from transformers import AutoTokenizer, BitsAndBytesConfig
import transformers
import torch

if torch.cuda.get_device_capability()[0] >= 8:
    !pip install -qqq flash-attn
    attn_implementation = "flash_attention_2"
    torch_dtype = torch.bfloat16
else:
    attn_implementation = "eager"
    torch_dtype = torch.float16

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch_dtype,
    bnb_4bit_use_double_quant=True,
)

model = "dfurman/Llama-3-70B-Orpo-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={
        "torch_dtype": torch_dtype,
        "quantization_config": bnb_config,
        "device_map": "auto",
        "attn_implementation": attn_implementation,
    }
)

Run

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Tell me a recipe for a spicy margarita."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print("***Prompt:\n", prompt)

outputs = pipeline(prompt, max_new_tokens=1000, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print("***Generation:\n", outputs[0]["generated_text"][len(prompt):])

Output

"""
"""

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	17.92
IFEval (0-Shot)	20.49
BBH (3-Shot)	24.09
MATH Lvl 5 (4-Shot)	13.52
GPQA (0-shot)	1.01
MuSR (0-shot)	16.28
MMLU-PRO (5-shot)	32.14