jpacifico/Chocolatine-14B-Instruct-DPO-v1.1

Chocolatine-14B-Instruct-DPO-v1.1

DPO fine-tuned of microsoft/Phi-3-medium-4k-instruct (14B params)
using the jpacifico/french-orca-dpo-pairs-revised rlhf dataset.
Training in French also improves the model in English, surpassing the performances of its base model.
Window context = 4k tokens

Benchmarks

The first Chocolatine-14B version is already the best-performing < 50B model in terms of MMLU-PRO on the OpenLLM Leaderboard (august 2024)
This new version 1.1 is also submitted, results coming soon.

MT-Bench

Chocolatine-14B-Instruct-DPO-v1.1 is outperforming Phi-3-medium-4k-instruct and its previous version.
And also this v1.1 is pretty close from GPT-4o-mini (first turn is amazing!).

########## First turn ##########
                                                     score
model                                         turn        
Chocolatine-14B-Instruct-DPO-v1.1             1     9.1375
gpt-4o-mini                                   1     9.1375
Chocolatine-14B-Instruct-4k-DPO               1     8.7250
Phi-3-medium-4k-instruct                      1     8.7125
Chocolatine-3B-Instruct-DPO-Revised           1     8.4625
Phi-3-mini-4k-instruct                        1     8.4125
gpt-3.5-turbo                                 1     8.2750

########## Second turn ##########
                                                      score
model                                         turn         
gpt-4o-mini                                   2     9.05000
gpt-3.5-turbo                                 2     8.20625
Chocolatine-14B-Instruct-DPO-v1.1             2     8.18750
Chocolatine-14B-Instruct-4k-DPO               2     8.15000
Phi-3-medium-4k-instruct                      2     7.92500
Chocolatine-3B-Instruct-DPO-Revised           2     7.61250
Phi-3-mini-4k-instruct                        2     7.38750

########## Average ##########
                                                  score
model                                                  
gpt-4o-mini                                    9.093750
Chocolatine-14B-Instruct-DPO-v1.1              8.662500
Chocolatine-14B-Instruct-4k-DPO                8.437500
Phi-3-medium-4k-instruct                       8.318750
gpt-3.5-turbo                                  8.240625
Chocolatine-3B-Instruct-DPO-Revised            8.037500
Phi-3-mini-4k-instruct                         7.900000

Usage

You can run this model using my Colab notebook

You can also run Chocolatine using the following code:

import transformers
from transformers import AutoTokenizer

# Format prompt
message = [
    {"role": "system", "content": "You are a helpful assistant chatbot."},
    {"role": "user", "content": "What is a Large Language Model?"}
]
tokenizer = AutoTokenizer.from_pretrained(new_model)
prompt = tokenizer.apply_chat_template(message, add_generation_prompt=True, tokenize=False)

# Create pipeline
pipeline = transformers.pipeline(
    "text-generation",
    model=new_model,
    tokenizer=tokenizer
)

# Generate text
sequences = pipeline(
    prompt,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    num_return_sequences=1,
    max_length=200,
)
print(sequences[0]['generated_text'])

Limitations

The Chocolatine model 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
• Model type: LLM
• Language(s) (NLP): French, English
• License: MIT