ProxyCoT SciTrek: Long-Context Scientific Reasoning Dataset

SciTrek data used in the ProxyCoT project, based on long-context reasoning tasks requiring 32,000 to 128,000 tokens. The dataset is hosted by 'oaimli' on Hugging Face and was last updated on July 7, 2026. It originates from a research paper published on arXiv.

Use Cases

• Benchmarking large language models on long-context scientific reasoning tasks based on the dataset's description.
• Training models for complex, multi-step inference requiring extended context windows up to 128K tokens.
• Researching Chain-of-Thought (CoT) or Proxy Chain-of-Thought methods for scientific problem-solving.

Strengths

• Explicitly designed for long-context reasoning tasks spanning 32K to 128K tokens.
• Associated with a named research project (ProxyCoT) and a cited arXiv paper (2509.21028).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Originally from the SciTrek research paper (arXiv:2509.21028).

Collection Method

Created for the ProxyCoT project; specific collection method not detailed.

Freshness

Last updated 2026-07-07 16:34:32; freshness should be verified.