Datasets:

fan-shu
/

swe-instruct-trajectories-empty-think-inserted

Modalities:

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Qwen3 Thinking SFT variant of fan-shu/swe-native-trajectories

Converted from fan-shu/swe-native-trajectories for supervised fine-tuning of Qwen3 Thinking models on trajectories without native CoT blocks.

Method A (ignore_empty_think): every assistant turn (text and/or tool calls) is wrapped with an empty <think>\n\n</think> prefix. Train with prime-rl mask_empty_think=true or ms-swift --loss_scale ignore_empty_think.

Recommended ms-swift loss scale: ignore_empty_think.

See Qwen3 ms-swift training docs.

Downloads last month: 18

Total file size: