Brokenarxiv Training Outputs Disprove: Qwen3.6-35B Responses to Perturbed ArXiv Statements

Training data generated from past ArXiv articles includes outputs from the Qwen3.6-35B model. The dataset contains the model's answers on whether perturbed mathematical statements are correct, with the expected answer always being disprove. It was created by MathArena and last updated on June 16, 2026.

Use Cases

• Evaluating LLM robustness against perturbed mathematical statements based on the described verification task.
• Training models for mathematical fact-checking using the generated Qwen3.6-35B outputs.
• Analyzing failure modes of language models on logical negation or perturbation of scientific text.

Strengths

• Data is derived from the established scientific preprint repository ArXiv.
• Model outputs are generated by a specific, named model (Qwen3.6-35B).
• The verification task has a clear, binary expected outcome (disprove).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

MathArena (https://matharena.ai/)

Collection Method

Generated from past ArXiv articles, with outputs produced by the Qwen3.6-35B model.

Freshness

Last updated 2026-06-16 14:17:34; freshness should be verified.