Datasets:

fan-shu
/

swe-native-trajectories-no-think

Modalities:

Formats:

Size:

Libraries:

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.

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

Method B (no_think): first user message gets /no_think; final assistant text uses the same empty think wrapper as method A.

Recommended ms-swift loss scale: default.