Why this matters: training student models to imitate advanced reasoning requires the teacher to expose full chain-of-thought (CoT), not a short summary. This dataset provides exactly that — runnable reasoning traces paired with final responses produced by DeepSeek-V4-Pro (reasoning_effort=max), making it a compact, ready-to-inspect distillation corpus for researchers experimenting with CoT supervision.
What Sets It Apart
- Full CoT as supervision: each example pairs a complete reasoning trace (reasoning_content) with the final response, rather than a summary. This makes the dataset directly usable for training students on “reasoning → response.”
- Teacher / prompt provenance: prompts are sampled from Jackrong/GLM-5.1-Reasoning-1M-Cleaned and a Math-focused split derived from MathForge, so examples cover mixed domains with a notable math subset.
- Compact, inspection-friendly format: ~1,000 samples (train + a train_math split), provided as JSONL and Parquet, with token-usage metadata per example to help cost/effort analysis.
Who It's For / Tradeoffs
Great fit if you want a small, high-quality corpus to prototype CoT distillation workflows (fine-tuning or student-teacher training) and to inspect raw reasoning traces for debugging model behavior. Look elsewhere if you need large-scale, curated, decontaminated benchmarks — this release is intentionally compact (~1K samples) and includes data derived from reformulations of MATH/MathForge, so there's a contamination risk when evaluating on those benchmarks.
Where It Fits
Use this dataset as a quick, low-cost experimental distillation source to: (1) prototype loss functions that supervise reasoning traces, (2) audit reasoning formats for downstream student models, or (3) build targeted math reasoning subsets. For large-scale production distillation, treat it as a pilot dataset and combine it with broader, decontaminated corpora.
Notes and cautions: licensed under Apache-2.0; teacher model labeled deepseek-v4-pro; the author is the Hugging Face user "beyoru". The dataset card flags potential contamination from MathForge / MATH-derived prompts — account for that in evaluation splits and downstream benchmarks.
