Large-model chain-of-thought traces have become a de facto training signal for improving LLM reasoning, but high‑quality, mixed‑model corpora are scattered and inconsistently formatted. This dataset aggregates and normalizes multi‑source CoT outputs into a single, sequence‑length capped corpus intended for supervised fine‑tuning and distilled reasoning training.
What Sets It Apart
- Multi‑model provenance: combines outputs from many public reasoning repositories and model families (DeepSeek‑v4, DeepSeek‑r1 variants, Qwen3/3.5/3.6, Gemma derivatives, and many distilled/synthetic reasoning sources). This breadth helps expose training pipelines to diverse reasoning styles rather than a single model’s biases.
- Preprocessed for SLM training: entries are normalized to ChatML-like fields and include repo_id, tok_len, user, thought_trace, assistant and ChatML to let you filter by source, token length, or user pattern before training or evaluation.
- Sequence-length constrained: content is curated to fit within a 5k token sequence window, making it practical for training models that accept long contexts while avoiding extremely long outliers.
- Composition transparency: the dataset documents per‑source token and row contributions (top contributors listed), enabling targeted exclusion or reweighting of large upstream corpora during dataset curation.
Who It's For and Trade-offs
Great fit if you want a ready-made, multi‑source CoT corpus to fine-tune or distill reasoning behavior into an SLM, prototype evaluation suites for step‑by‑step solutions, or perform experiments that compare reasoning styles across model families. The dataset is released with Apache‑2.0 licensing metadata on the card, and includes tooling‑friendly formats (JSON) and column metadata for selective filtering.
Look elsewhere if you require purely human‑annotated reasoning chains, provenance guarantees for each example beyond repository attribution, or complete removal of synthetic/automatically generated content: this corpus intentionally aggregates model outputs and synthetic examples (and thus inherits their noise and hallucination risk). Also plan for downstream validation and filtering if you need high precision or safety‑critical correctness.
