Reasoning Corpus 5M (reasoning-corpus-4K-5M-v1)

Summary

Aggregated corpus of model-generated reasoning chains (chain-of-thought) filtered to within a ~5k-token sequence length, mixed from 93 upstream Hugging Face source splits. Largest contributors by tokens: glaiveai/reasoning-v1-20m (19.52%), PrimeIntellect/INTELLECT-3-SFT openreasoning_science (9.87%) and am_chat (8.28%), BAAI/OpenSeek-Synthetic-Reasoning-Data-Examples CC (6.34%), nvidia/Nemotron-Cascade-SFT-Stage-1 general (4.93%), open-thoughts/OpenThoughts2-1M (4.78%). Per the card the traces originate from DeepSeek-v4/DeepSeek-R1, Qwen3 / Qwen3.5-3.6 and Gemma4-31B family models. Each row keeps repo_id, tok_len, user, thought_trace, assistant and a preformatted ChatML field, so the same data can be re-templated for any target model's chat format. Shipped as a single dataset.jsonl; the card recommends the HF streaming interface instead of a full download, and states tok_len is an estimate and the traces are model-generated training data rather than verified proofs.

Keywords

hf-dataset reasoning cot chain-of-thought sft distillation text-generation agentic code jsonl chatml streaming machine-generated deepseek-v4 qwen3

Topics

Language Modeling / Reasoning

Research notes

• downloads=10108; likes=200 (checked 2026-08-11). Card inconsistency: the load example sets DATASET_ID to 'SupraLabs/reasoning-corpus-4K-5M-v1', not the live Qyrou/ path; the footer says the work was created by @QyrouNnet-AI and transferred to this org at SupraLabs' request. Row count not independently verifiable - the HF viewer conversion is partial (412,596 converted, ~3.67M estimated), so the ~5M figure rests on the card's own source table. Card explicitly declines to guarantee 100% accuracy and recommends dedup, source balancing and trace filtering before full fine-tuning. No linked paper, GitHub repo or eval results.