Deepseek V4 Pro Max Distill 1500X: Reasoning Traces for Model Distillation

A dataset containing reasoning traces and final answers generated by the DeepSeek-V4-Pro model with maximum reasoning effort. The dataset was created by the author 'beyoru' and was last updated on May 4, 2026. It consists of 1000 samples generated from prompts sourced from the Jackrong/GLM-5.1-Reasoning-1M-Cleaned dataset.

Use Cases

• Training smaller models via knowledge distillation based on the reasoning traces.
• Analyzing the reasoning patterns of a large language model based on the generated traces.
• Evaluating the quality of generated reasoning steps for model benchmarking.
• Developing new prompting strategies based on the sampled prompts from the source dataset.

Strengths

• Contains 1000 samples, providing a defined scale for analysis.
• Generated by a state-of-the-art model (DeepSeek-V4-Pro) with reasoning explicitly enabled.
• Cost information is provided, indicating a low generation cost of approximately $5.46.
• Prompts are sourced from a known, cleaned reasoning dataset, suggesting a structured origin.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is limited to 1000 samples, which may be insufficient for large-scale training.
• The dataset's primary stated goal is for quality checking, suggesting it may not be intended for broad, direct application.

Provenance

Source

Generated by DeepSeek-V4-Pro model using prompts from Jackrong/GLM-5.1-Reasoning-1M-Cleaned.

Collection Method

Model inference with reasoning_effort=max, thinking.enabled=true.

Time Range

Dataset created around April 27, 2026.

Freshness

Last updated 2026-05-04 11:28:06; freshness should be verified.

Geography

null