ShareGPT-Qwen3-4B-T0.6-Thinking-Regen

# ShareGPT — Qwen3-4B Thinking Regeneration (T0.6, 32k budget) This dataset contains 36,315 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`. The prompt set (system and user messages, including the "You are a helpful assistant." system turn) is exactly the one of [TY233/ShareGPT-Qwen3-4B-T0.7-NonThinking-Regen](https://huggingface.co/datasets/TY233/ShareGPT-Qwen3-4B-T0.7-NonThinking-Regen) (36,936 conversations), so the two datasets form a thinking/non-thinking pair over shared prompts. ## Generation | Parameter | Value | | --- | --- | | Model | Qwen/Qwen3-4B | | Model revision | `1cfa9a7208912126459214e8b04321603b3df60c` | | Thinking | Enabled | | Temperature | 0.6 | | Top-p | 0.95 | | Top-k | 20 | | Min-p | 0.0 | | Maximum new tokens per assistant turn | 32,768 | Sampling follows the Qwen3 model card's recommended settings for thinking mode. The 32,768-token budget leaves reasoning traces uncapped in practice (about 0.2% of turns reach it), so the dataset reflects the model's full thinking-length distribution. Relation to [TY233/ShareGPT-Qwen3-4B-T0.7-Thinking-Regen](https://huggingface.co/datasets/TY233/ShareGPT-Qwen3-4B-T0.7-Thinking-Regen): same model and prompt source, but that dataset uses the non-thinking sampling settings (temperature 0.7, top-p 0.8) and a 4,096-token generation budget, which truncates and filters out longer reasoning traces. This dataset uses the thinking-mode recommended sampling and an effectively unconstrained budget. ## Filtering Of the 36,936 input conversations, 36,315 (98.3%) were regenerated successfully and are released here. A conversation is retained only if **every assistant turn** has non-empty `content` and non-empty `reasoning_content` and neither field contains residual `<think>`/`</think>` markers. The 621 dropped conversations: 371 exceeded the model's 40,960-token context window with the 32k generation budget, 228 produced an empty field in at least one turn (mostly generations that hit the token budget mid-reasoning), 21 request timeouts, and 1 malformed input row. Retained conversations follow the input order. Conversation identifiers are unique; user and system messages are unchanged from the inputs. ## 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. ## Token-length statistics (Qwen3-4B tokenizer) Full conversation (prompt + all turns incl. reasoning), from a 495-conversation pilot with identical settings: p50 ≈ 4.1k, p90 ≈ 10.4k, p95 ≈ 13.4k, p99 ≈ 20.4k tokens. Thinking roughly doubles sequence lengths relative to the non-thinking counterpart.