Tokenskip Math Qwen3 4B Thinking N8: Compressed Chain-of-Thought Outputs for MATH

Intermediate artifacts for reproducing the TokenSkip method on the MATH dataset. The data consists of chain-of-thought outputs generated by the Qwen3-4B-Thinking-2507 model, subsequently compressed with LLMLingua-2. The dataset was created by user 'leapeto' and uploaded to Hugging Face on April 28, 2026.

Use Cases

• Preparing supervised fine-tuning data for language models based on pre-generated and compressed reasoning traces.
• Studying the effects of text compression on chain-of-thought reasoning quality.
• Reproducing or building upon the TokenSkip pipeline for mathematical problem-solving.
• Analyzing the output diversity of a language model on math questions, based on the N=8 samples per question.

Strengths

• Provides pre-computed outputs from a specific model (Qwen3-4B-Thinking-2507) for the MATH dataset, saving computational resources.
• Includes compressed versions of the reasoning traces, which may reduce storage and training costs.
• Contains N=8 samples per question, allowing for analysis of generation diversity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is an intermediate artifact for a specific pipeline; its standalone utility may be limited.

Provenance

Source

Hugging Face dataset uploaded by user 'leapeto'.

Collection Method

Generated using vLLM on 4× GPUs, with N=8 samples per question at temperature T=0.7 and a maximum generation length of 4,096 tokens, then compressed with LLMLingua-2.

Time Range

The generation date is not specified, but the dataset was uploaded on April 28,我们发现了一个错误。

Freshness

Last updated 2026-04-28 04:12:38; freshness should be verified.

Geography

The dataset is not inherently geographic.