CoolFace
Datasetpublic

JingweiNi/ocr2_cf1900_k2_qwen35_gpt55_aggregated_10k_seed20260513

Aggregated Qwen3.5 and GPT-5.5 OCR2 Coding Step Labels This dataset follows the same 500-row parquet layout as JingweiNi/ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513 and aggregates the original Qwen3.5 labels with GPT-5.5 medium reannotation of Qwen3.5-positive error steps. Aggregation rule: GPT-5.5 confirmed errors are errors; all other 10,000 manifest-selected steps are correct. Columns qwen35_verified: original Qwen3.5 labels for the 10,000 manifest-selected… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ocr2_cf1900_k2_qwen35_gpt55_aggregated_10k_seed20260513.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face
0likes16downloads
Dataset Card

Aggregated Qwen3.5 and GPT-5.5 OCR2 Coding Step Labels

This dataset follows the same 500-row parquet layout as JingweiNi/ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513 and aggregates the original Qwen3.5 labels with GPT-5.5 medium reannotation of Qwen3.5-positive error steps.

Aggregation rule: GPT-5.5 confirmed errors are errors; all other 10,000 manifest-selected steps are correct.

Columns

  • —qwen35_verified: original Qwen3.5 labels for the 10,000 manifest-selected steps, NaN elsewhere.
  • —gpt55_verified: GPT-5.5 labels for the 1,536 Qwen3.5-positive steps, NaN elsewhere.
  • —verified: aggregated labels, with 1.0 only for GPT-5.5 confirmed errors, 0.0 for all other 10,000 manifest-selected steps, and NaN outside the manifest.

Aggregate Counts

  • —Aggregate valid labels: 10,000
  • —Aggregate errors: 730
  • —Aggregate correct labels: 9,270
  • —Non-manifest binary labels: 0

Metadata files under metadata/ include the source 10K manifest, selected Qwen-positive manifest, GPT-5.5 run summary, and OpenRouter usage summary.