jaeyong2/Qwen2.5-3B-Instruct-Ja-SFT

This model was trained on a Japanese dataset and built with Qwen.

Evaluation

llm-jp-eval script(colab)

!git clone https://github.com/llm-jp/llm-jp-eval.git
!cd llm-jp-eval && pip install -e .
!cd llm-jp-eval && python scripts/preprocess_dataset.py --dataset-name all --output-dir ./dataset_dir
!cd llm-jp-eval && python scripts/evaluate_llm.py -cn config.yaml model.pretrained_model_name_or_path=jaeyong2/Qwen2.5-0.5B-Instruct-JaMagpie-Preview tokenizer.pretrained_model_name_or_path=jaeyong2/Qwen2.5-0.5B-Instruct-JaMagpie-Preview dataset_dir=./dataset_dir/1.4.1/evaluation/test

llm-jp-eval	Qwen2.5-3B-Instruct	finetuning-model
AVG	0.4921	0.4895
CG	0.1000	0
EL	0.4770	0.4431
FA	0.1210	0.1246
HE	0.5550	0.5650
MC	0.7133	0.7900
MR	0.5400	0.6100
MT	0.6391	0.5982
NLI	0.6640	0.6640
QA	0.2638	0.3165
RC	0.8481	0.7837

License

Qwen/Qwen2.5-3B-Instruct : https://huggingface.co./Qwen/Qwen2.5-3B-Instruct/blob/main/LICENSE

Acknowledgement

This research is supported by TPU Research Cloud program.