TIE-Pilot/qwen3-4b-perfectblend-deepspec-rollout

Qwen3-4B PerfectBlend DeepSpec Rollout

This dataset contains the complete DeepSpec-aligned Qwen3-4B self-distillation rollout over the filtered PerfectBlend corpus. The seeded 95/5 split is published as separate train and eval splits.

Splits

Data construction

• —Canonical filtered corpus: 1,420,906 conversations.
• —Split: seed 42, 95% train / 5% eval.
• —The original assistant messages are discarded. Every assistant turn is regenerated in order, and each generated answer is included in the context for subsequent turns in the same conversation.
• —Generator: Qwen/Qwen3-4B, with thinking disabled, for both splits.
• —Sampling: temperature 0.7, top-p 0.8, top-k 20, min-p 0.
• —Maximum generated tokens per assistant turn: 4,096.
• —No deduplication is applied.
• —Canonical source content SHA256: 9d1376be67ae4c4708145d6939c4d8d49e467e5c4b78773fce633b81ea256613.
• —Reconstructed eval input SHA256: 291ba4429bf74e9dc3541cb9679a119bf924fa6edd4c9494d5f2bd523f97441d.

Schema

Each JSONL row is one conversation:

{ "id": 123, "conversations": [ {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."} ], "status": "success" }

The id is the global ID from the canonical filtered dataset and is used for exact-ID resume and coverage validation.

Integrity

The published snapshot was validated before upload:

• —Train: 128/128 shards, 1,349,860 rows, and 1,349,860 unique IDs.
• —Eval: 64/64 shards, 71,046 rows, and 71,046 unique IDs.
• —Eval generated IDs exactly equal the seeded source eval IDs.
• —Train/eval ID intersection is zero; their union is exactly 1,420,906 IDs.
• —No malformed JSON, failed statuses, duplicate IDs, missing IDs, or empty assistant responses were found.
• —Train sorted global-ID SHA256: 625a89eada47d80063a25f31c920bca23272abfb86e6a433094b36a93be899be.
• —Remote file counts and byte sizes were checked after upload; generated file contents are retained in the validated ucsd-hopper snapshot.

Four records contain a whitespace-only user turn inherited unchanged from the canonical source data.

Use of this generated dataset remains subject to the licenses and terms of the upstream PerfectBlend data and Qwen3-4B model.

Citation

This dataset is the training corpus behind Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting (arXiv:2608.27339), and behind the drafters released at `TIE-Pilot/dspark-attnconv-block7-qwen3-4b` and `TIE-Pilot/deepspec-drafter-ablations`.

@misc{qiang2026parallelblindnessinformationfloors,
      title={Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting}, 
      author={Xinwei Qiang and Xiang Fang and Chang Chen and Zaifeng Pan and Yue Guan and Yufei Ding},
      year={2026},
      eprint={2608.27339},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.27339}, 
}