mgoin/open-perfectblend-glm5.2-regen

open-perfectblend-glm5.2-regen

On-policy regeneration of the full mlabonne/open-perfectblend with GLM-5.2-FP8, built to train speculative-decoding drafters (dspark / DFlash).

• —1,420,229 conversations in ShareGPT-style {id, conversations: [{from, value}], source}.
• —Each assistant turn was regenerated by GLM-5.2-FP8 conditioned on the preceding, already-regenerated turns — deepspec-style per-turn on-policy regeneration, up to 8k tokens per turn. Original human turns are preserved.
• —Assistant turns contain the model's inline `<think>…</think>` reasoning followed by the answer.
• —The source column marks each row's origin within the blend: open-perfectblend (1,212,473, the blend minus its ultrachat subset) and ultrachat (207,756, regenerated in a separate batch). Together they cover the whole blend.

Generation method

• —Target model: GLM-5.2-FP8, served with vLLM.
• —For each user turn, the full prior context (with previously-regenerated assistant turns) is sent to the model and its response replaces the original assistant turn, so every assistant turn is on-policy for the target model.

Intended use

On-policy training data for speculative-decoding draft models that accelerate GLM-5.2.

Provenance & license

Derived from mlabonne/open-perfectblend, which aggregates multiple source datasets — refer to the source dataset for licensing of the underlying prompts. Assistant responses were regenerated by GLM-5.2.