OCR2 Hardest10K: Aggregated Step Labels for AI Reasoning Traces

10,000 reasoning traces from the hardest OCR2 questions, aggregated from multiple AI models. The dataset was created by JingweiNi and last updated on May 30, 2026. Each row contains step-level labels from Qwen3.5-122B and GPT-5.5 models, formatted as aligned arrays.

Use Cases

• Training or fine-tuning models for step-by-step reasoning based on the aggregated step labels.
• Benchmarking model verification capabilities using the Qwen3.5 and GPT-5.5 step annotations.
• Analyzing error patterns in AI reasoning on difficult questions from the OCR2 benchmark.

Strengths

• Contains 10,000 reasoning traces from the hardest subset of the OCR2 benchmark.
• Provides aggregated step-level labels from two distinct AI models (Qwen3.5-122B and GPT-5.5).
• Follows a specific, documented row-level aggregated format from a related dataset.

Limitations

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

Provenance

Source

huggingface

Collection Method

Aggregated labels from AI model (Qwen3.5-122B, GPT-5.5) annotations on generated reasoning traces.

Freshness

Last updated 2026-05-30 18:04:39.