Improve language tag
#1
by
lbourdois
- opened
README.md
CHANGED
|
@@ -1,164 +1,175 @@
|
|
| 1 |
-
---
|
| 2 |
-
library_name: transformers
|
| 3 |
-
license: apache-2.0
|
| 4 |
-
datasets:
|
| 5 |
-
- HoangHa/Pensez-v0.1
|
| 6 |
-
language:
|
| 7 |
-
-
|
| 8 |
-
-
|
| 9 |
-
|
| 10 |
-
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
##
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
| Pensez
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
|
| 63 |
-
|
| 64 |
-
|
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
|
| 87 |
-
|
| 88 |
-
|
|
| 89 |
-
|
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
- [Modal](https://modal.com/)
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
datasets:
|
| 5 |
+
- HoangHa/Pensez-v0.1
|
| 6 |
+
language:
|
| 7 |
+
- zho
|
| 8 |
+
- eng
|
| 9 |
+
- fra
|
| 10 |
+
- spa
|
| 11 |
+
- por
|
| 12 |
+
- deu
|
| 13 |
+
- ita
|
| 14 |
+
- rus
|
| 15 |
+
- jpn
|
| 16 |
+
- kor
|
| 17 |
+
- vie
|
| 18 |
+
- tha
|
| 19 |
+
- ara
|
| 20 |
+
base_model:
|
| 21 |
+
- Qwen/Qwen2.5-7B-Instruct
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
<div align="center">
|
| 25 |
+
|
| 26 |
+
# Pensez: Less Data, Better Reasoning – Rethinking French LLM
|
| 27 |
+
|
| 28 |
+
[**About**](#about) | [**How to Run Locally**](#run-locally) | [**Models and Datasets**](#models-and-datasets) | [**Benchmarks**](#benchmarks) | [**Training Details**](#training-details)
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
</div>
|
| 32 |
+
|
| 33 |
+
## About
|
| 34 |
+
|
| 35 |
+
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.
|
| 36 |
+
|
| 37 |
+
Key strategies for improved reasoning:
|
| 38 |
+
- **Concise reasoning** for simple tasks to prevent overthinking.
|
| 39 |
+
- **Extended reasoning** for complex domains like mathematics, coding, and science.
|
| 40 |
+
- **Special tokens (`<think>...</think>`)** to explicitly guide the model’s reasoning process.
|
| 41 |
+
|
| 42 |
+
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).
|
| 43 |
+
|
| 44 |
+
## Models and Datasets
|
| 45 |
+
|
| 46 |
+
### Model Versions
|
| 47 |
+
|
| 48 |
+
Pensez is built upon [Qwen 2.5 Instruct 7B](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) and trained over five epochs.
|
| 49 |
+
|
| 50 |
+
| Model | Backbone | Size | Download Link |
|
| 51 |
+
|---------------|----------------------------------------|------|---------------|
|
| 52 |
+
| Pensez-v0.1-e1 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e1](https://huggingface.co/HoangHa/Pensez-v0.1-e1) |
|
| 53 |
+
| Pensez-v0.1-e2 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e2](https://huggingface.co/HoangHa/Pensez-v0.1-e2) |
|
| 54 |
+
| Pensez-v0.1-e3 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e3](https://huggingface.co/HoangHa/Pensez-v0.1-e3) |
|
| 55 |
+
| Pensez-v0.1-e4 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e4](https://huggingface.co/HoangHa/Pensez-v0.1-e4) |
|
| 56 |
+
| Pensez-v0.1-e5 | Qwen2.5-7B-Instruct | 7B | [🤗 Pensez-v0.1-e5](https://huggingface.co/HoangHa/Pensez-v0.1-e5) |
|
| 57 |
+
|
| 58 |
+
### Dataset
|
| 59 |
+
|
| 60 |
+
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).
|
| 61 |
+
|
| 62 |
+
| Dataset | Description | Size | Link |
|
| 63 |
+
|--------------|----------------------|-------|-------|
|
| 64 |
+
| Pensez v0.1 | SFT Training Dataset | 2K samples | [🤗 Pensez v0.1](https://huggingface.co/datasets/HoangHa/Pensez-v0.1) |
|
| 65 |
+
|
| 66 |
+
## Benchmarks
|
| 67 |
+
|
| 68 |
+
Pensez was evaluated on French-specific benchmarks, demonstrating strong reasoning ability and improved task-specific performance:
|
| 69 |
+
|
| 70 |
+
| Benchmark | Pensez-v0.1-e5 | DeepSeek-R1-Distill-Qwen-7B | Qwen2.5-7B-Instruct |
|
| 71 |
+
|-----------|---------------|-----------------------------|----------------------|
|
| 72 |
+
| Math-hard (fr) | 0.3458 | 0.3403 | 0.2253 |
|
| 73 |
+
| MMLU (fr) | 0.5766 | 0.4961 | 0.6612 |
|
| 74 |
+
| BoolQA (fr) | 0.9157 | 0.7079 | 0.9382 |
|
| 75 |
+
| Trivia (en) | 0.4421 | 0.2711 | 0.5316 |
|
| 76 |
+
| HellaSwag (en) | 0.5050 | 0.3540 | 0.5258 |
|
| 77 |
+
|
| 78 |
+
**Key Observations:**
|
| 79 |
+
- Pensez outperforms Qwen2.5-7B-Instruct in reasoning tasks.
|
| 80 |
+
- Comparable to DeepSeek-R1-Distill-Qwen-7B in reasoning while maintaining strong understanding.
|
| 81 |
+
- Reduced degradation in knowledge-based tasks.
|
| 82 |
+
|
| 83 |
+
<details>
|
| 84 |
+
<summary>Click for detailed benchmark results</summary>
|
| 85 |
+
|
| 86 |
+
| 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 |
|
| 87 |
+
|------------------------------------------------|---------------|---------------|---------------|---------------|---------------|-----------------|-----------|
|
| 88 |
+
| leaderboard_math_hard_fr | 0.0918 | 0.2547 | 0.2783 | 0.3035 | 0.3458 | 0.2253 | 0.3403 |
|
| 89 |
+
| leaderboard_math_algebra_hard_fr | 0.1029 | 0.3914 | 0.3971 | 0.5114 | 0.5000 | 0.4229 | 0.4771 |
|
| 90 |
+
| leaderboard_math_counting_and_prob_hard_fr | 0.0765 | 0.1378 | 0.1939 | 0.2041 | 0.2398 | 0.1224 | 0.2347 |
|
| 91 |
+
| leaderboard_math_geometry_hard_fr | 0.0388 | 0.1019 | 0.1408 | 0.1359 | 0.1748 | 0.1019 | 0.2330 |
|
| 92 |
+
| leaderboard_math_num_theory_hard_fr | 0.1198 | 0.2581 | 0.3502 | 0.3548 | 0.4332 | 0.3180 | 0.3963 |
|
| 93 |
+
| leaderboard_math_prealgebra_hard_fr | 0.1681 | 0.4425 | 0.4690 | 0.4956 | 0.5841 | 0.3274 | 0.4867 |
|
| 94 |
+
| leaderboard_math_precalculus_hard_fr | 0.0357 | 0.0714 | 0.1190 | 0.1190 | 0.1429 | 0.0595 | 0.2143 |
|
| 95 |
+
| leaderboard_mmlu_fr | 0.3806 | 0.3329 | - | - | 0.5766 | 0.6612 | 0.4961 |
|
| 96 |
+
| french_bench_arc_challenge | 0.5047 | 0.5021 | 0.4919 | 0.4859 | 0.4842 | 0.5518 | 0.3447 |
|
| 97 |
+
| french_bench_boolqa | 0.9326 | 0.9326 | 0.9326 | 0.9270 | 0.9157 | 0.9382 | 0.7079 |
|
| 98 |
+
| french_bench_fquadv2 | 0.4325 | 0.4400 | 0.4412 | 0.4375 | 0.4387 | 0.4800 | 0.2988 |
|
| 99 |
+
| french_bench_hellaswag | 0.4970 | 0.5055 | 0.5092 | 0.5058 | 0.5050 | 0.5258 | 0.3540 |
|
| 100 |
+
| french_bench_trivia | 0.4763 | 0.4763 | 0.4553 | 0.4395 | 0.4421 | 0.5316 | 0.2711 |
|
| 101 |
+
|
| 102 |
+
</details>
|
| 103 |
+
|
| 104 |
+
## Run Locally
|
| 105 |
+
|
| 106 |
+
You can run Pensez using Hugging Face’s `transformers` library:
|
| 107 |
+
|
| 108 |
+
```python
|
| 109 |
+
import torch
|
| 110 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 111 |
+
|
| 112 |
+
model_path = "HoangHa/Pensez-v0.1-e5"
|
| 113 |
+
|
| 114 |
+
# Load tokenizer and model
|
| 115 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
| 116 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 117 |
+
model_path, torch_dtype=torch.float16, device_map="auto"
|
| 118 |
+
)
|
| 119 |
+
|
| 120 |
+
# Example input
|
| 121 |
+
messages = [{"role": "user", "content": "Bonjour!"}]
|
| 122 |
+
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors='pt').to("cuda")
|
| 123 |
+
|
| 124 |
+
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)
|
| 125 |
+
response = tokenizer.decode(generated_ids[0], skip_special_tokens=True, clean_up_tokenization_space=True)
|
| 126 |
+
print(f"Réponse: {response}")
|
| 127 |
+
```
|
| 128 |
+
|
| 129 |
+
## Training Details
|
| 130 |
+
|
| 131 |
+
Pensez was trained with:
|
| 132 |
+
- **Packing Inputs Without Cross-Contamination Attention** ([Reference](https://github.com/MeetKai/functionary/tree/main/functionary/train/packing))
|
| 133 |
+
- **Liger Kernel** ([Reference](https://github.com/linkedin/Liger-Kernel))
|
| 134 |
+
- **DeepSpeed 3** ([Reference](https://github.com/deepspeedai/DeepSpeed))
|
| 135 |
+
- **NEFTune Noise** ([Reference](https://arxiv.org/abs/2310.05914)) for robustness.
|
| 136 |
+
|
| 137 |
+
| **Parameter** | **Value** |
|
| 138 |
+
|--------------|----------|
|
| 139 |
+
| Epochs | 5 |
|
| 140 |
+
| Global Batch Size | 200 |
|
| 141 |
+
| Learning Rate | 1e-5 |
|
| 142 |
+
| Scheduler | Cosine |
|
| 143 |
+
| Optimizer | AdamW |
|
| 144 |
+
| Warmup Ratio | 0.05 |
|
| 145 |
+
| Weight Decay | 0.01 |
|
| 146 |
+
| Max Sequence Length | 16,384 |
|
| 147 |
+
|
| 148 |
+
More details: [Training Config]() | Loss curves: [Wandb](https://wandb.ai/hahuyhoanghhh41/llamafactory?nw=nwuserhahuyhoanghhh41)
|
| 149 |
+
|
| 150 |
+
## Citation
|
| 151 |
+
|
| 152 |
+
```bibtex
|
| 153 |
+
@misc{dao2025alphamazeenhancinglargelanguage,
|
| 154 |
+
title={Pensez: Less Data, Better Reasoning – Rethinking French LLM},
|
| 155 |
+
author={Ha Huy Hoang},
|
| 156 |
+
year={2025},
|
| 157 |
+
archivePrefix={arXiv},
|
| 158 |
+
primaryClass={cs.CL},
|
| 159 |
+
url={},
|
| 160 |
+
}
|
| 161 |
+
```
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
## Acknowledgement
|
| 165 |
+
|
| 166 |
+
- [llama-factory](https://github.com/hiyouga/LLaMA-Factory)
|
| 167 |
+
- [Deepseek R1](https://github.com/deepseek-ai/DeepSeek-R1)
|
| 168 |
+
- [Qwen 2.5](https://github.com/QwenLM/Qwen2.5)
|
| 169 |
+
- [NEFTune Noise](https://arxiv.org/abs/2310.05914)
|
| 170 |
+
- [Packing Inputs Without Cross-Contamination Attention](https://github.com/MeetKai/functionary/tree/main/functionary/train/packing)
|
| 171 |
+
- [Liger Kernel](https://github.com/linkedin/Liger-Kernel)
|
| 172 |
+
- [Deepspeed](https://github.com/deepspeedai/DeepSpeed)
|
| 173 |
+
- [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)
|
| 174 |
+
- [Hyperbolic](https://hyperbolic.xyz/)
|
| 175 |
- [Modal](https://modal.com/)
|