90 lines
3.1 KiB
Markdown
90 lines
3.1 KiB
Markdown
|
---
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- small-lm
|
||
|
|
- math
|
||
|
|
- reasoning
|
||
|
|
- slm
|
||
|
|
- french
|
||
|
|
license: apache-2.0
|
||
|
|
datasets:
|
||
|
|
- openai/gsm8k
|
||
|
|
- kurakurai/luth-sft
|
||
|
|
- cmh/gsm8k_fr
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen3-0.6B
|
||
|
|
language:
|
||
|
|
- fr
|
||
|
|
- en
|
||
|
|
---
|
||
|
|
|
||
|
|
# Qwen3-0.6B-Fr
|
||
|
|
|
||
|
|
This model is obtained by fine-tuning Qwen/Qwen3-0.6B on the [kurakurai/luth-sft](https://huggingface.co/datasets/kurakurai/luth-sft) dataset, specifically
|
||
|
|
subsets luth_smoltalk2, luth_aya_dataset, luth_croissantllm and luth_tulu3_persona_instruct.
|
||
|
|
The model is used in the experiments described in https://bknyaz.github.io/blog/2026/meta-merge/.
|
||
|
|
Single A100 was used for fine-tuning and evaluation.
|
||
|
|
|
||
|
|
The following versions were used for train/eval:
|
||
|
|
|
||
|
|
- python >= 3.10
|
||
|
|
- torch : 2.9.0+cu128
|
||
|
|
- lm_eval : 0.4.9.1
|
||
|
|
- vllm : 0.11.1
|
||
|
|
- transformers : 4.57.6
|
||
|
|
- datasets : 3.2.0
|
||
|
|
- numpy : 2.2.6
|
||
|
|
|
||
|
|
## Training
|
||
|
|
|
||
|
|
The [TRL](https://github.com/huggingface/trl) library was used with SFT/full-rank options:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
python trl/scripts/sft.py --model_name_or_path Qwen/Qwen3-0.6B --dataset_name kurakurai/luth-sft --dataset_config main --learning_rate 2e-5 \
|
||
|
|
--num_train_epochs 1 --per_device_train_batch_size 2 --gradient_accumulation_steps 8 --gradient_checkpointing --eos_token '<|im_end|>' --eval_strategy no \
|
||
|
|
--completion_only_loss True --report_to wandb --output_dir /path/to/the/finetuned/model
|
||
|
|
```
|
||
|
|
|
||
|
|
This is by far not the most compute and performance efficient fine-tuning, but it could be a good baseline.
|
||
|
|
|
||
|
|
The dataset was preprocessed to the conversational format:
|
||
|
|
|
||
|
|
```python
|
||
|
|
# trl/scripts/sft.py
|
||
|
|
|
||
|
|
dataset = load_dataset(...)
|
||
|
|
|
||
|
|
trainer = SFTTrainer(
|
||
|
|
model=model,
|
||
|
|
args=training_args,
|
||
|
|
train_dataset=concatenate_datasets([dataset['luth_smoltalk2'], dataset['luth_aya_dataset'], dataset['luth_croissantllm'], dataset['luth_tulu3_persona_instruct']]),
|
||
|
|
eval_dataset=dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None,
|
||
|
|
peft_config=get_peft_config(model_args),
|
||
|
|
)
|
||
|
|
|
||
|
|
```
|
||
|
|
|
||
|
|
## Evaluation
|
||
|
|
|
||
|
|
Evaluation was done with lm_eval on the test split of [gsm8k](https://huggingface.co/datasets/openai/gsm8k),
|
||
|
|
[french_bench (avg score)](https://github.com/EleutherAI/lm-evaluation-harness/tree/main/lm_eval/tasks/french_bench) and [gsm8k-fr](https://huggingface.co/datasets/cmh/gsm8k_fr):
|
||
|
|
|
||
|
|
```bash
|
||
|
|
python -m lm_eval --model vllm --model_args pretrained=${model},tensor_parallel_size=1,dtype=auto,gpu_memory_utilization=0.9,data_parallel_size=1 \
|
||
|
|
--tasks gsm8k,french_bench,gsm8k-fr --batch_size 1 --apply_chat_template=True --confirm_run_unsafe_code --trust_remote_code
|
||
|
|
```
|
||
|
|
|
||
|
|
To evaluate on gsm8k-fr you can use our fork https://github.com/bknyaz/lm-evaluation-harness/tree/main/lm_eval/tasks/gsm8k.
|
||
|
|
|
||
|
|
### Results
|
||
|
|
|
||
|
|
| Model | gsm8k| french | gsm8k-fr | avg |
|
||
|
|
|-----------------------|------|--------|----------|------|
|
||
|
|
| Qwen3-0.6B | 21.0 | 24.4 | 19.6 | 21.7 |
|
||
|
|
| Qwen3-0.6B-Fr | 36.1 | 26.5 | 26.5 | 29.7 |
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
## License
|
||
|
|
|
||
|
|
Please refer to the license of the original model [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) and dataset [kurakurai/luth-sft](https://huggingface.co/datasets/kurakurai/luth-sft).
|