CoolFace
Modelpublic

HoangHa/Pensez-v0.1-e1

sourceHugging Faceapache-2.0updated 2y agoView on Hugging Face
0likes14downloads
Model Card

<div align="center">

Pensez: Less Data, Better Reasoning – Rethinking French LLM

**About** | **How to Run Locally** | **Models and Datasets** | **Benchmarks** | **Training Details**

image/png </div>

About

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.

Models and Datasets

Model Versions

Pensez is built upon Qwen 2.5 Instruct 7B and trained over five epochs.

ModelBackboneSizeDownload Link
Pensez-v0.1-e1Qwen2.5-7B-Instruct7B🤗 Pensez-v0.1-e1
Pensez-v0.1-e2Qwen2.5-7B-Instruct7B🤗 Pensez-v0.1-e2
Pensez-v0.1-e3Qwen2.5-7B-Instruct7B🤗 Pensez-v0.1-e3
Pensez-v0.1-e4Qwen2.5-7B-Instruct7B🤗 Pensez-v0.1-e4
Pensez-v0.1-e5Qwen2.5-7B-Instruct7B🤗 Pensez-v0.1-e5

Dataset

Pensez was trained on the hand-curated Pensez v0.1 dataset containing 2,000 samples (1,000 French, 1,000 English).

DatasetDescriptionSizeLink
Pensez v0.1SFT Training Dataset2K samples🤗 Pensez v0.1

Benchmarks

Pensez was evaluated on French-specific benchmarks, demonstrating strong reasoning ability and improved task-specific performance:

BenchmarkPensez-v0.1-e5DeepSeek-R1-Distill-Qwen-7BQwen2.5-7B-Instruct
Math-hard (fr)0.34580.34030.2253
MMLU (fr)0.57660.49610.6612
BoolQA (fr)0.91570.70790.9382
Trivia (en)0.44210.27110.5316
HellaSwag (en)0.50500.35400.5258

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>

TasksPensez v0.1 e1Pensez v0.1 e2Pensez v0.1 e3Pensez v0.1 e4Pensez v0.1 e5Qwen 7B instructR1 distil
leaderboardmathhard_fr0.09180.25470.27830.30350.34580.22530.3403
leaderboardmathalgebrahardfr0.10290.39140.39710.51140.50000.42290.4771
leaderboardmathcountingandprobhardfr0.07650.13780.19390.20410.23980.12240.2347
leaderboardmathgeometryhardfr0.03880.10190.14080.13590.17480.10190.2330
leaderboardmathnumtheoryhard_fr0.11980.25810.35020.35480.43320.31800.3963
leaderboardmathprealgebrahardfr0.16810.44250.46900.49560.58410.32740.4867
leaderboardmathprecalculushardfr0.03570.07140.11900.11900.14290.05950.2143
leaderboardmmlufr0.38060.3329--0.57660.66120.4961
frenchbencharc_challenge0.50470.50210.49190.48590.48420.55180.3447
frenchbenchboolqa0.93260.93260.93260.92700.91570.93820.7079
frenchbenchfquadv20.43250.44000.44120.43750.43870.48000.2988
frenchbenchhellaswag0.49700.50550.50920.50580.50500.52580.3540
frenchbenchtrivia0.47630.47630.45530.43950.44210.53160.2711

</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:

**Parameter****Value**
Epochs5
Global Batch Size200
Learning Rate1e-5
SchedulerCosine
OptimizerAdamW
Warmup Ratio0.05
Weight Decay0.01
Max Sequence Length16,384

More details: [Training Config]() | Loss curves: Wandb

Citation

bibtex
@misc{dao2025alphamazeenhancinglargelanguage,
      title={Pensez: Less Data, Better Reasoning – Rethinking French LLM},
      author={Ha Huy Hoang},
      year={2025},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={},
}

Acknowledgement