dvlab-research/Step-DPO

share sft-dataset

yyht opened this issue · 4 comments

yyht commented

hello, nice work. could share the sft-dataset in hf?

Sure, it will be released soon. Please stay tuned.

Hi authors, following up on this thread to stay updated when the SFT datasets are released. Thanks and nice work!

Hi authors, it is a nice work to advance the off-policy method for enhancing reason ability of LLM. I am following up on this thread to stay updated when the SFT datasets are released. Thanks!

yyht commented

hello everyone, https://huggingface.co/datasets/yingyingzhang/metamath-qwen2-math .
I use qwen2-math-instruct and open-source-datasets such as metamath-qa and numina-cot to construct a high quality sft-dataset.
When finetuning on qwen2-general-base or qwen2-math-base, the sft model could achieve comparable results to qwen2-instruct-7b\72b and qwen2-math-7b-instruct.
The whole datasets contains metamath-qwen2-math and none-synthetic datasets from https://huggingface.co/datasets/AI-MO/NuminaMath-CoT.
Please enjoy it.