DPO fine-tuning of the merged model jpacifico/Chocolatine-2-merged-qwen25arch (Qwen-2.5-14B architecture)
using the jpacifico/french-orca-dpo-pairs-revised RLHF dataset.
Training in French also improves the model's overall capabilities.

[!TIP] Window context : up to 128K tokens

LLM Leaderboard FR

[Updated 2025-10-20]

Top 3 all categories on the French Government Leaderboard LLM FR

COLE Benchmark (Laval University)

[Updated 2025-10-20]

Chocolatine-2-14B-Instruct-v2.0.3 stands out as the strongest open-weights models with a Composite Score of 45.05%
Paper here with benchmark results

MT-Bench-French

Chocolatine-2 outperforms its previous versions and its base architecture Qwen-2.5 model on MT-Bench-French, used with multilingual-mt-bench and GPT-4-Turbo as a LLM-judge.
My goal was to achieve GPT-4o-mini's performance on the French language, this version comes close to the performance of the OpenAI model according to this benchmark

########## First turn ##########
                                             score
model                                 turn        
gpt-4o-mini                           1     9.287500
Chocolatine-2-14B-Instruct-v2.0.3     1     9.112500
Qwen2.5-14B-Instruct                  1     8.887500
Chocolatine-14B-Instruct-DPO-v1.2     1     8.612500
Phi-3.5-mini-instruct                 1     8.525000
Chocolatine-3B-Instruct-DPO-v1.2      1     8.375000
DeepSeek-R1-Distill-Qwen-14B          1     8.375000
phi-4                                 1     8.300000
Phi-3-medium-4k-instruct              1     8.225000
gpt-3.5-turbo                         1     8.137500
Chocolatine-3B-Instruct-DPO-Revised   1     7.987500
Meta-Llama-3.1-8B-Instruct            1     7.050000
vigostral-7b-chat                     1     6.787500
Mistral-7B-Instruct-v0.3              1     6.750000
gemma-2-2b-it                         1     6.450000

########## Second turn ##########
                                               score
model                                 turn
Chocolatine-2-14B-Instruct-v2.0.3     2     9.050000         
gpt-4o-mini                           2     8.912500
Qwen2.5-14B-Instruct                  2     8.912500
Chocolatine-14B-Instruct-DPO-v1.2     2     8.337500
DeepSeek-R1-Distill-Qwen-14B          2     8.200000
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
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

########## Average ##########
                                          score
model                                          
gpt-4o-mini                            9.100000
Chocolatine-2-14B-Instruct-v2.0.3      9.081250
Qwen2.5-14B-Instruct                   8.900000
Chocolatine-14B-Instruct-DPO-v1.2      8.475000
DeepSeek-R1-Distill-Qwen-14B           8.287500
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
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

OpenLLM Leaderboard (Archived)

Chocolatine-2 is the best-performing 14B fine-tuned model (Ex-aequo with avg. score 41.08) on the OpenLLM Leaderboard
[Updated 2025-02-12]

Metric	Value
Avg.	41.08
IFEval	70.37
BBH	50.63
MATH Lvl 5	40.56
GPQA	17.23
MuSR	19.07
MMLU-PRO	48.60

Usage

You can run this model using my Colab notebook

You can also run Chocolatine-2 using the following code:

python

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: Apache-2.0

Made with ❤️ in France

jpacifico/Chocolatine-2-14B-Instruct-v2.0.3

Chocolatine-2-14B-Instruct-v2.0.3