metadata
base_model:
- Qwen/Qwen2.5-7B-Instruct
datasets:
- HoangHa/Pensez-v0.1
language:
- en
- fr
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
<div align="center">
# Pensez: Less Data, Better Reasoning – Rethinking French LLM
[**About**](#about) | [**How to Run Locally**](#run-locally) | [**Models and Datasets**](#models-and-datasets) | [**Benchmarks**](#benchmarks) | [**Training Details**](#training-details)
</div>
## About
Paper: [Pensez: Less Data, Better Reasoning - Rethinking French LLM](https://huggingface.co/papers/2503.13661)
Pensez is a bilingual (French-English) reasoning model designed to maximize efficiency with significantly reduced training data. The model leverages a curated dataset focusing on daily reasoning tasks and scientific questions to enhance performance.
Key strategies for improved reasoning:
- **Concise reasoning** for simple tasks to prevent overthinking.
- **Extended reasoning** for complex domains like mathematics, coding, and science.
- **Special tokens (`<think>...</think>`)** to explicitly guide the model’s reasoning process.
These optimizations result in superior reasoning capabilities while maintaining robust general understanding compared to models like [DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B).
## Models and Datasets
### Model Versions
Pensez is built upon [Qwen 2.5 Instruct 7B](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) and trained over five epochs.
| Model | Backbone | Size | Download Link |
|---------------|----------------------------------------|------|---------------|
| Pensez-v0.1-e1 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e1](https://huggingface.co/HoangHa/Pensez-v0.1-e1) |
| Pensez-v0.1-e2 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e2](https://huggingface.co/HoangHa/Pensez-v0.1-e2) |
| Pensez-v0.1-e3 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e3](https://huggingface.co/HoangHa/Pensez-v0.1-e3) |
| Pensez-v0.1-e4 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e4](https://huggingface.co/HoangHa/Pensez-v0.1-e4) |
| Pensez-v0.1-e5 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e5](https://huggingface.co/HoangHa/Pensez-v0.1-e5) |
### Dataset
Pensez was trained on the hand-curated [Pensez v0.1](https://huggingface.co/datasets/HoangHa/Pensez-v0.1) dataset containing 2,000 samples (1,000 French, 1,000 English).
| Dataset | Description | Size | Link |
|--------------|----------------------|-------|-------|
| Pensez v0.1 | SFT Training Dataset | 2K samples | [🤗 Pensez v0.1](https://huggingface.co/datasets/HoangHa/Pensez-v0.1) |
## Benchmarks
Pensez was evaluated on French-specific benchmarks, demonstrating strong reasoning ability and improved task-specific performance:
| Benchmark | Pensez-v0.1-e5 | DeepSeek-R1-Distill-Qwen-7B | Qwen2.5-7B-Instruct |
|-----------|---------------|-----------------------------|----------------------|
| Math-hard (fr) | 0.3458 | 0.3403 | 0.2253 |
| MMLU (fr) | 0.5766 | 0.4961 | 0.6612 |
| BoolQA (fr) | 0.9157 | 0.7079 | 0.9382 |
| Trivia (en) | 0.4421 | 0.2711 | 0.5316 |
| HellaSwag (en) | 0.5050 | 0.3540 | 0.5258 |
**Key Observations:**
- Pensez outperforms Qwen2.5-7B-Instruct in reasoning tasks.
- Comparable to DeepSeek-R1-Distill-Qwen-7B in reasoning while maintaining strong understanding.
- Reduced degradation in knowledge-based tasks.
<details>
<summary>Click for detailed benchmark results</summary>
| Tasks | Pensez v0.1 e1 | Pensez v0.1 e2 | Pensez v0.1 e3 | Pensez v0.1 e4 | Pensez v0.1 e5 | Qwen 7B instruct | R1 distil |
|------------------------------------------------|---------------|---------------|---------------|---------------|---------------|-----------------|-----------|
| leaderboard_math_hard_fr | 0.0918 | 0.2547 | 0.2783 | 0.3035 | 0.3458 | 0.2253 | 0.3403 |
| leaderboard_math_algebra_hard_fr | 0.1029 | 0.3914 | 0.3971 | 0.5114 | 0.5000 | 0.4229 | 0.4771 |
| leaderboard_math_counting_and_prob_hard_fr | 0.0765 | 0.1378 | 0.1939 | 0.2041 | 0.2398 | 0.1224 | 0.2347 |
| leaderboard_math_geometry_hard_fr | 0.0388 | 0.1019 | 0.1408 | 0.1359 | 0.1748 | 0.1019 | 0.2330 |
| leaderboard_math_num_theory_hard_fr | 0.1198 | 0.2581 | 0.3502 | 0.3548 | 0.4332 | 0.3180 | 0.3963 |
| leaderboard_math_prealgebra_hard_fr | 0.1681 | 0.4425 | 0.4690 | 0.4956 | 0.5841 | 0.3274 | 0.4867 |
| leaderboard_math_precalculus_hard_fr | 0.0357 | 0.0714 | 0.1190 | 0.1190 | 0.1429 | 0.0595 | 0.2143 |
| leaderboard_mmlu_fr | 0.3806 | 0.3329 | - | - | 0.5766 | 0.6612 | 0.4961 |
| french_bench_arc_challenge | 0.5047 | 0.5021 | 0.4919 | 0.4859 | 0.4842 | 0.5518 | 0.3447 |
| french_bench_boolqa | 0.9326 | 0.9326 | 0.9326 | 0.9270 | 0.9157 | 0.9382 | 0.7079 |
| french_bench_fquadv2 | 0.4325 | 0.4400 | 0.4412 | 0.4375 | 0.4387 | 0.4800 | 0.2988 |
| french_bench_hellaswag | 0.4970 | 0.5055 | 0.5092 | 0.5058 | 0.5050 | 0.5258 | 0.3540 |
| french_bench_trivia | 0.4763 | 0.4763 | 0.4553 | 0.4395 | 0.4421 | 0.5316 | 0.2711 |
</details>
## Run Locally
You can run Pensez using Hugging Face’s `transformers` library:
```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_path = "HoangHa/Pensez-v0.1-e5"
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
model_path, torch_dtype=torch.float16, device_map="auto"
)
# Example input
messages = [{"role": "user", "content": "Bonjour!"}]
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors='pt').to("cuda")
generated_ids = model.generate(input_ids, max_new_tokens=2500, temperature=0.8, repetition_penalty=1.1, do_sample=True, eos_token_id=tokenizer.eos_token_id)
response = tokenizer.decode(generated_ids[0], skip_special_tokens=True, clean_up_tokenization_space=True)
print(f"Réponse: {response}")
Training Details
Pensez was trained with:
- Packing Inputs Without Cross-Contamination Attention (Reference)
- Liger Kernel (Reference)
- DeepSpeed 3 (Reference)
- NEFTune Noise (Reference) for robustness.
| Parameter | Value |
|---|---|
| Epochs | 5 |
| Global Batch Size | 200 |
| Learning Rate | 1e-5 |
| Scheduler | Cosine |
| Optimizer | AdamW |
| Warmup Ratio | 0.05 |
| Weight Decay | 0.01 |
| Max Sequence Length | 16,384 |
More details: Training Config | Loss curves: Wandb
Citation
@misc{ha2025pensezreasoningfrenchllm,
title={Pensez: Less Data, Better Reasoning – Rethinking French LLM},
author={Ha Huy Hoang},
year={2025},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2503.13661},
}
Acknowledgement
- llama-factory
- Deepseek R1
- Qwen 2.5
- NEFTune Noise
- Packing Inputs Without Cross-Contamination Attention
- Liger Kernel
- Deepspeed
- lm-evaluation-harness
- Hyperbolic
- Modal