Chocolatine-3B-Instruct-DPO-Revised

Chocolatine-3B-Instruct-DPO-Revised

DPO fine-tuned of microsoft/Phi-3-mini-4k-instruct (3.82B 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

Quantized 4-bit and 8-bit versions are available (see below)
A larger version Chocolatine-14B is also available in its latest version-1.2

Benchmarks

Chocolatine is the best-performing 3B model on the OpenLLM Leaderboard (august 2024)
[Update 2024-08-22] Chocolatine-3B also outperforms Microsoft's new model Phi-3.5-mini-instruct on the average benchmarks of the 3B category.

MT-Bench-French

Chocolatine-3B-Instruct-DPO-Revised is outperforming GPT-3.5-Turbo on MT-Bench-French, used with multilingual-mt-bench and GPT-4-Turbo as LLM-judge.
Notably, this latest version of the Chocolatine-3B model is approaching the performance of Phi-3-Medium (14B) in French.

########## First turn ##########
                                                      score
model                                         turn         
gpt-4o-mini                                   1     9.28750
Chocolatine-14B-Instruct-DPO-v1.2             1     8.61250
Phi-3-medium-4k-instruct                      1     8.22500
gpt-3.5-turbo                                 1     8.13750
Chocolatine-3B-Instruct-DPO-Revised           1     7.98750
Daredevil-8B                                  1     7.88750
NeuralDaredevil-8B-abliterated                1     7.62500
Phi-3-mini-4k-instruct                        1     7.21250
Meta-Llama-3.1-8B-Instruct                    1     7.05000
vigostral-7b-chat                             1     6.78750
Mistral-7B-Instruct-v0.3                      1     6.75000
gemma-2-2b-it                                 1     6.45000
French-Alpaca-7B-Instruct_beta                1     5.68750
vigogne-2-7b-chat                             1     5.66250

########## Second turn ##########
                                                       score
model                                         turn          
gpt-4o-mini                                   2     8.912500
Chocolatine-14B-Instruct-DPO-v1.2             2     8.337500
Chocolatine-3B-Instruct-DPO-Revised           2     7.937500
Phi-3-medium-4k-instruct                      2     7.750000
gpt-3.5-turbo                                 2     7.679167
NeuralDaredevil-8B-abliterated                2     7.125000
Daredevil-8B                                  2     7.087500
Meta-Llama-3.1-8B-Instruct                    2     6.787500
Mistral-7B-Instruct-v0.3                      2     6.500000
Phi-3-mini-4k-instruct                        2     6.487500
vigostral-7b-chat                             2     6.162500
gemma-2-2b-it                                 2     6.100000
French-Alpaca-7B-Instruct_beta                2     5.487395
vigogne-2-7b-chat                             2     2.775000

########## Average ##########
                                                  score
model                                                  
gpt-4o-mini                                    9.100000
Chocolatine-14B-Instruct-DPO-v1.2              8.475000
Phi-3-medium-4k-instruct                       7.987500
Chocolatine-3B-Instruct-DPO-Revised            7.962500
gpt-3.5-turbo                                  7.908333
Daredevil-8B                                   7.487500
NeuralDaredevil-8B-abliterated                 7.375000
Meta-Llama-3.1-8B-Instruct                     6.918750
Phi-3-mini-4k-instruct                         6.850000
Mistral-7B-Instruct-v0.3                       6.625000
vigostral-7b-chat                              6.475000
gemma-2-2b-it                                  6.275000
French-Alpaca-7B-Instruct_beta                 5.587866
vigogne-2-7b-chat                              4.218750

Quantized versions

• 4-bit quantized version is available here : jpacifico/Chocolatine-3B-Instruct-DPO-Revised-Q4_K_M-GGUF

• 8-bit quantized version also available here : jpacifico/Chocolatine-3B-Instruct-DPO-Revised-Q8_0-GGUF

• Ollama: jpacifico/chocolatine-3b

ollama run jpacifico/chocolatine-3b

Ollama Modelfile example :

FROM ./chocolatine-3b-instruct-dpo-revised-q4_k_m.gguf
TEMPLATE """{{ if .System }}<|system|>
{{ .System }}<|end|>
{{ end }}{{ if .Prompt }}<|user|>
{{ .Prompt }}<|end|>
{{ end }}<|assistant|>
{{ .Response }}<|end|>
"""
PARAMETER stop """{"stop": ["<|end|>","<|user|>","<|assistant|>"]}"""
SYSTEM """You are a friendly assistant called Chocolatine."""

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