Datasetpublic

Huang2020/qwen3.6-27B-reasoning-regen

Qwen3.6-27B Reasoning Regen Successful ShareGPT and PerfectBlend conversations regenerated with a local Qwen3.6-27B checkpoint. No exact public checkpoint revision was recorded for the run. Config Source Rows sharegpt_full Aeala/ShareGPT_Vicuna_unfiltered 78,753 sharegpt_exploded sharegpt_full 233,443 perfectblend_full mlabonne/open-perfectblend 1,419,275 perfectblend_exploded perfectblend_full 1,882,975 The *_full configs contain successful regenerated… See the full description on the dataset page: https://huggingface.co/datasets/Huang2020/qwen3.6-27B-reasoning-regen.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

3likes134downloads

Dataset Card

Qwen3.6-27B Reasoning Regen

Successful ShareGPT and PerfectBlend conversations regenerated with a local Qwen3.6-27B checkpoint. No exact public checkpoint revision was recorded for the run.

Config	Source	Rows
`sharegpt_full`	`Aeala/ShareGPT_Vicuna_unfiltered`	78,753
`sharegpt_exploded`	`sharegpt_full`	233,443
`perfectblend_full`	`mlabonne/open-perfectblend`	1,419,275
`perfectblend_exploded`	`perfectblend_full`	1,882,975

The *_full configs contain successful regenerated conversations. The *_exploded configs are deterministic post-processing with one row per assistant generation event; they are not another generation run. Historical assistant reasoning_content is removed in exploded rows, while the final target assistant keeps it.

Generation configuration

yaml

target_model: Qwen3.6-27B  # local checkpoint; revision not recorded
api: /v1/chat/completions
enable_thinking: true
reasoning_parser: qwen3
reasoning: save
temperature: 0
max_tokens: 32768
top_p: not sent  # model default was 0.95
top_k: not sent  # model default was 20

For multi-turn rows, all source assistant responses were regenerated turn by turn. Previously generated visible assistant content was used as history for later turns, but historical hidden reasoning_content was removed from the next request.

The machine-readable version is `generation_config.json`.

Schema

Full rows contain id, conversations, and status. Assistant messages have role, content, and structured reasoning_content.

Exploded rows additionally contain source_id, source_row_index, and assistant_turn_index. source_row_index is the zero-based row in the corresponding successful full JSONL, not the original Hugging Face source row.

Only status="success" rows are included. Skipped/error sidecars and runtime artifacts are not uploaded. The generator did not persist or reject by finish_reason, so a response reaching the 32,768-token limit can still be marked successful.

License

This repository does not assert additional rights over source conversations or model outputs. Users must follow the terms of the two source datasets and the Qwen3.6-27B checkpoint.