500 reasoning traces and final answers generated by DeepSeek-V4-Pro (reasoning_effort=max, thinking.enabled=true). The prompts were sampled from the first 500 rows of the train split of Jackrong/GLM-5.1-Reasoning-1M-Cleaned dataset. The dataset was created by beyoru and last updated on Hugging Face in April 2026.
Use Cases
- Training smaller reasoning models based on the detailed reasoning traces.
- Evaluating the quality and consistency of a large model's reasoning process.
- Analyzing the relationship between prompt complexity and generated reasoning steps.
- Creating benchmark datasets for model distillation techniques.
Strengths
- Contains 500 samples, providing a substantial preview set.
- Traces are generated by a state-of-the-art model (DeepSeek-V4-Pro) with maximum reasoning effort enabled.
- Prompts are sourced from a known, cleaned reasoning dataset (GLM-5.1-Reasoning-1M-Cleaned).
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is limited to 500 samples, which may not be sufficient for full-scale training.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- Hugging Face dataset created by beyoru.
- Collection Method
- Generated by DeepSeek-V4-Pro model processing prompts from another dataset.
- Freshness
- Last updated 2026-04-24 06:50:40; freshness should be verified.