Chocolatine-14B-Instruct-DPO-v1.3
DPO fine-tuning of microsoft/Phi-4 (14B params)
using the jpacifico/french-orca-dpo-pairs-revised rlhf dataset.
Training in French also improves the model's overall capabilities, surpassing the performances of its base model.
Window context = up to 16k tokens
OpenLLM Leaderboard
Could this be the biggest performance boost ever seen from LLM fine-tuning ? 🤔
Chocolatine-14B-Instruct-DPO-v1.3 is the best-performing Phi-4 based model on the OpenLLM Leaderboard
for only 1.70kgCo2 (versus > 3kg for other models in the same category and performance)
[Updated 2025-02-17]
| Metric | Value |
|---|---|
| Avg. | 42.42 |
| IFEval | 70.40 |
| BBH | 54.85 |
| MATH Lvl 5 | 56.19 |
| GPQA | 12.19 |
| MuSR | 12.29 |
| MMLU-PRO | 48.60 |
MT-Bench-French
Chocolatine-14B-Instruct-DPO-v1.3 outperforms its previous Chocolatine versions and its base model Phi-4 on MT-Bench-French, used with multilingual-mt-bench and GPT-4-Turbo as LLM-judge.
########## First turn ##########
score
model turn
gpt-4o-mini 1 9.2875
Chocolatine-14B-Instruct-DPO-v1.3 1 9.0125
Chocolatine-14B-Instruct-DPO-v1.2 1 8.6125
Phi-3.5-mini-instruct 1 8.5250
Chocolatine-3B-Instruct-DPO-v1.2 1 8.3750
phi-4 1 8.3000
Phi-3-medium-4k-instruct 1 8.2250
gpt-3.5-turbo 1 8.1375
Chocolatine-3B-Instruct-DPO-Revised 1 7.9875
Daredevil-8B 1 7.8875
Meta-Llama-3.1-8B-Instruct 1 7.0500
vigostral-7b-chat 1 6.7875
Mistral-7B-Instruct-v0.3 1 6.7500
gemma-2-2b-it 1 6.4500
French-Alpaca-7B-Instruct_beta 1 5.6875
vigogne-2-7b-chat 1 5.6625
########## Second turn ##########
score
model turn
gpt-4o-mini 2 8.912500
Chocolatine-14B-Instruct-DPO-v1.3 2 8.762500
Chocolatine-14B-Instruct-DPO-v1.2 2 8.337500
phi-4 2 8.131250
Chocolatine-3B-Instruct-DPO-Revised 2 7.937500
Chocolatine-3B-Instruct-DPO-v1.2 2 7.862500
Phi-3-medium-4k-instruct 2 7.750000
gpt-3.5-turbo 2 7.679167
Phi-3.5-mini-instruct 2 7.575000
Daredevil-8B 2 7.087500
Meta-Llama-3.1-8B-Instruct 2 6.787500
Mistral-7B-Instruct-v0.3 2 6.500000
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-v1.3 8.825000
Chocolatine-14B-Instruct-DPO-v1.2 8.475000
phi-4 8.215625
Chocolatine-3B-Instruct-DPO-v1.2 8.118750
Phi-3.5-mini-instruct 8.050000
Phi-3-medium-4k-instruct 7.987500
Chocolatine-3B-Instruct-DPO-Revised 7.962500
gpt-3.5-turbo 7.908333
Daredevil-8B 7.487500
Meta-Llama-3.1-8B-Instruct 6.918750
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
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-2 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, 2025
- Model type: LLM
- Language(s) (NLP): French, English
- License: MIT
Made with ❤️ in France