CoolFace
Datasetpublic

TY233/ShareGPT-Qwen3-4B-T0.7-Thinking-Regen

ShareGPT — Qwen3-4B Thinking Regeneration This dataset contains 33,590 ShareGPT conversations with assistant responses regenerated by Qwen/Qwen3-4B in thinking mode. User prompts are retained; each regenerated assistant turn includes its reasoning in reasoning_content and its final answer in content. Generation Parameter Value Model Qwen/Qwen3-4B Model revision 1cfa9a7208912126459214e8b04321603b3df60c Thinking Enabled Temperature 0.7 Top-p 0.8… See the full description on the dataset page: https://huggingface.co/datasets/TY233/ShareGPT-Qwen3-4B-T0.7-Thinking-Regen.

sourceHugging Faceupdated 7d agoView on Hugging Face
0likes38downloads
Dataset Card

ShareGPT — Qwen3-4B Thinking Regeneration

This dataset contains 33,590 ShareGPT conversations with assistant responses regenerated by Qwen/Qwen3-4B in thinking mode. User prompts are retained; each regenerated assistant turn includes its reasoning in reasoning_content and its final answer in content.

Generation

ParameterValue
ModelQwen/Qwen3-4B
Model revision1cfa9a7208912126459214e8b04321603b3df60c
ThinkingEnabled
Temperature0.7
Top-p0.8
Top-k20
Min-p0.0
Maximum new tokens per assistant turn4,096

Filtering

We retain successfully regenerated conversations only when every assistant turn contains both a non-empty final answer and non-empty reasoning. Responses with residual <think> or </think> markers inside either field are filtered out. Conversation identifiers are checked for uniqueness, and user/system messages are checked against the generation inputs. Retained conversations follow the input order.

The prompt cohort was selected from the earlier non-thinking regeneration: conversations were formatted with the Qwen chat template, tokenized with a 3,072-token limit, and required to retain at least two adjacent supervised assistant tokens. This is a prompt-cohort selection rule, not a truncation of the thinking responses released here.

The released JSONL retains the generated text; the 3,072-token capture/training limit is not applied to this file.

Format

data/train.jsonl contains one conversation per line, with an id and a conversations list. Each message has role and content; assistant messages additionally contain reasoning_content.

These are model-generated responses, not verified reasoning traces or correctness annotations.