jpacifico/Chocolatine-Cook-3B-combined-SFT-DPO-v0.1

Description model

Chocolatine-3B version specialized in French culinary language, fine-tuning of microsoft/Phi-3.5-mini-instruct.
This model is based on 283 specific terms and definitions of French cuisine.

Fine Tuning

For this version of the model I experimented a training method with a double fine-tuning, SFT then DPO.
I generated two datasets exclusively for this model, with GPT-4o deployed on Azure OpenAI.
The challenge was to achieve a consistent alignment between the two fine-tuning methods.
SFT to teach the terms and DPO to reinforce the understanding achieved during the first learning.

Fine tuning done efficiently with Unsloth, with which I saved processing time on a single T4 GPU (AzureML compute instance).

Usage

The recommended usage is by loading the low-rank adapter using unsloth:

from unsloth import FastLanguageModel
from transformers import TextStreamer
import torch

model_name = "jpacifico/Chocolatine-Cook-3B-combined-SFT-DPO-v0.1"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name,
    max_seq_length=2048,
    dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    load_in_4bit=False
)

FastLanguageModel.for_inference(model)
model.eval()

def generate_response(user_question: str):
    messages = [
        {"role": "system", "content": "Tu es un assistant IA spécialisé dans le langage culinaire français. Une question te sera posée. Tu dois générer une réponse précise et concise."},
        {"role": "user", "content": "En cuisine "+user_question},
    ]

    inputs = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt",
    ).to("cuda")

    attention_mask = (inputs != tokenizer.pad_token_id).long()

    text_streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

    with torch.no_grad():
        _ = model.generate(
            input_ids=inputs,
            attention_mask=attention_mask,
            max_new_tokens=128,
            use_cache=True,
            streamer=text_streamer,
            do_sample=False,
            temperature=0.7,
        )

Limitations

The Chocolatine model series is a quick demonstration that a base model can be easily fine-tuned to achieve compelling performance.
It does not have any moderation mechanism.

• Developed by: Jonathan Pacifico, 2024
• License: MIT
• Finetuned from model : microsoft/Phi-3.5-mini-instruct