95 lines
4.1 KiB
Markdown
95 lines
4.1 KiB
Markdown
---
|
|
library_name: transformers
|
|
datasets:
|
|
- kurakurai/luth-sft
|
|
language:
|
|
- fr
|
|
- en
|
|
base_model:
|
|
- LiquidAI/LFM2-350M
|
|
pipeline_tag: text-generation
|
|
license: other
|
|
license_name: lfm1.0
|
|
license_link: LICENSE
|
|
tags:
|
|
- liquid
|
|
- lfm2
|
|
- luth
|
|
---
|
|
|
|
|
|
|
|
|
|
# Luth-LFM2-350M
|
|
|
|
**Luth-LFM2-350M** is a French fine-tuned version of [LFM2-350M](https://huggingface.co/LiquidAI/LFM2-350M) in collaboration with Liquid AI, trained on the [Luth-SFT](https://huggingface.co/datasets/kurakurai/luth-sft) dataset. The model has improved its French capabilities in instruction following, math, and general knowledge. Additionally, its English capabilities have remained stable.
|
|
|
|
Our Evaluation, training and data scripts are available on [GitHub](https://github.com/kurakurai/Luth), along with the [Blog](https://huggingface.co/blog/MaxLSB/luth) we wrote, to further detail our recipe.
|
|
|
|
|
|
|
|
|
|
|
|
## Model Details
|
|
|
|
The model was trained using full fine-tuning on the Luth-SFT dataset with [Axolotl](https://github.com/axolotl-ai-cloud/axolotl). The resulting model was then merged back with LFM2-350M. This process successfully retained the model's English capabilities while improving its performance in French.
|
|
|
|
## Benchmark Results
|
|
|
|
We used LightEval for evaluation, with custom tasks for the French benchmarks. The models were evaluated with a `temperature=0`.
|
|
|
|
### French Benchmark Scores
|
|
|
|
| Model | IFEval<br>French | GPQA-Diamond<br>French | MMLU<br>French | Math500<br>French | Arc-Challenge<br>French | Hellaswag<br>French |
|
|
| --------------------- | ------------------ | ------------------------ | ---------------- | ------------------- | ------------------------- | --------------------- |
|
|
| **Luth-LFM2-350M** | <u>38.26</u> | 26.40 | <u>39.15</u> | <u>23.00</u> | <u>34.13</u> | <u>43.39</u> |
|
|
| LFM2-350M | 31.55 | <u>28.93</u> | 38.63 | 18.00 | 33.36 | 39.13 |
|
|
| SmolLM2-360M-Instruct | 21.50 | 28.43 | 26.14 | 3.20 | 26.60 | 32.94 |
|
|
|
|
|
|
### English Benchmark Scores
|
|
|
|
| Model | IFEval <br> English | GPQA-Diamond <br> English | MMLU <br> English | Math500 <br> English | Arc-Challenge <br> English | Hellaswag <br> English |
|
|
| --------------------- | ------------------- | ------------------------- | ----------------- | -------------------- | -------------------------- | ---------------------- |
|
|
| **Luth-LFM2-350M** | <u>57.05</u> | <u>28.28</u> | 44.36 | <u>23.20</u> | 34.81 | 45.92 |
|
|
| LFM2-350M | 56.81 | 27.27 | <u>44.79</u> | 20.87 | 34.27 | 45.07 |
|
|
| SmolLM2-360M-Instruct | 33.95 | 20.71 | 26.18 | 3.00 | <u>35.41</u> | <u>52.17</u> |
|
|
|
|
|
|
## Code Example
|
|
|
|
```python
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained("kurakurai/Luth-LFM2-350M")
|
|
model = AutoModelForCausalLM.from_pretrained("kurakurai/Luth-LFM2-350M")
|
|
messages = [
|
|
{"role": "user", "content": "Quelle est la capitale de la France?"},
|
|
]
|
|
inputs = tokenizer.apply_chat_template(
|
|
messages,
|
|
add_generation_prompt=True,
|
|
tokenize=True,
|
|
return_dict=True,
|
|
return_tensors="pt",
|
|
).to(model.device)
|
|
|
|
outputs = model.generate(**inputs, max_new_tokens=100)
|
|
print(
|
|
tokenizer.decode(
|
|
outputs[0][inputs["input_ids"].shape[-1] :], skip_special_tokens=True
|
|
)
|
|
)
|
|
```
|
|
|
|
## Citation
|
|
|
|
```bibtex
|
|
@misc{luth2025kurakurai,
|
|
title = {Luth: Efficient French Specialization for Small Language Models and Cross-Lingual Transfer},
|
|
author = {Lasbordes, Maxence and Gad, Sinoué},
|
|
year = {2025},
|
|
howpublished = {\url{https://arxiv.org/abs/2510.05846}},
|
|
note = {arXiv:2510.05846}
|
|
}
|
|
``` |