README.md

---
library_name: transformers
license: apache-2.0
datasets:
- kurakurai/luth-sft
language:
- fr
- en
base_model:
- Qwen/Qwen3-1.7B
pipeline_tag: text-generation
---

![Kurakura AI Logo](media/logo_kurakura.png)


# Luth-1.7B-Instruct

**Luth-1.7B-Instruct** is a French fine-tuned version of [Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B), trained on the [Luth-SFT](https://huggingface.co/datasets/kurakurai/luth-sft) dataset. The model has drastically improved its French capabilities in instruction following, math, and general knowledge. Additionally, its English capabilities have remained stable and have even increased in some areas.

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.

![Luth Graph](media/luth-graph.png)

## Model Details

Luth 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 with the base Qwen3-1.7B model. This process successfully retained the model's English capabilities while improving its performance on most selected benchmarks in both French and English.

## 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-1.7B-Instruct** | <u>58.53</u>       | <u>36.55</u>             | <u>49.75</u>      | <u>62.60</u>       | 35.16                  | 31.88             |
| Qwen3-1.7B             | 54.71           | 31.98                 | 28.49          | 60.40           | 33.28                  | 24.86             |
| SmolLM2-1.7B-Instruct  | 30.93           | 20.30                 | 33.73          | 10.20           | 28.57                  | <u>49.58</u>         |
| Qwen2.5-1.5B-Instruct  | 31.30           | 27.41                 | 46.25          | 33.20           | 32.68                  | 34.33             |
| LFM2-1.2B              | 54.41           | 22.84                 | 47.59          | 36.80           | <u>39.44</u>              | 33.05             |

### 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-1.7B-Instruct** | 65.80           | 29.80                  | <u>60.28</u>      | 70.40            | 42.24                   | 58.53              |
| Qwen3-1.7B             | <u>68.88</u>       | <u>31.82</u>              | 52.82          | <u>71.20</u>        | 36.18                   | 46.98              |
| SmolLM2-1.7B-Instruct  | 49.04           | 25.08                  | 50.27          | 22.67            | 42.32                   | <u>66.94</u>          |
| Qwen2.5-1.5B-Instruct  | 39.99           | 25.76                  | 59.81          | 57.20            | 41.04                   | 64.48            |
| LFM2-1.2B              | 68.52           | 24.24                  | 55.22          | 45.80            | <u>42.58</u>               | 57.61              |

## Code Example

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("kurakurai/Luth-1.7B-Instruct")
model = AutoModelForCausalLM.from_pretrained("kurakurai/Luth-1.7B-Instruct")
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}
}
```