Rename Pro splits, case IDs, document paths, and Harbor tasks

#7
by xujustinj - opened

Use the final six challenge names in split files, case IDs, PDF paths, category tags, card links, and Harbor metadata.

Original split Final split Public name
document_conventions interpretation_rules Interpretation rules
house_vocabulary custom_terminology Custom terminology
house_counting_rules conditional_aggregation Conditional aggregation
drawn_relations relational_diagrams Relational diagrams
charts_inside_documents charts_and_maps Charts and maps
long_chart_sheets long_single_pages Long single pages

The names describe the primary task or page format. Split membership is unchanged. All 100 case-ID prefixes use the final namespace; the migration map records each original ID/path and unchanged PDF hash.

Harbor task names derive from the document identifier, independently of the challenge. All 100 document-based task names are unchanged from the preceding review candidate. Source split, case ID, and tags remain in metadata. Original published Harbor names have no aliases.

Validation:

  • All 100 original PDF hashes, schemas, expected outputs, field rules, and repeated-structure payloads preserved.
  • All 100 task execution payloads preserved outside identity metadata and challenge tags; task configs pin the final data revision and unchanged scorer.
  • Final PR tree contains exactly six final split files and 100 task directories; superseded review split/PDF paths are removed.
  • Existing deterministic reference scoring passed 100/100 with Value F1 1.0; prior Docker oracle/nop rewards were 1.0/0.0, with zero exceptions. Execution inputs and code remain unchanged.
  • Matching unpublished Hub candidate has 100 validated configurations and content-pinned hashes.
  • HF main remains at 60b101709fe6e4c2925a494ef98e91f9ecbc300c.

Final data revision: d2882e554029ce9dda8b1f50eb471df26ff7f72c. Full candidate: 1b0778e0625c55e80e1d001a2963d5b081963a9c. Scorer: 2f27f100730afc7c395602e26a2a87e07ea66b31.

The framework translates original HF layouts and group names. Pro downloads and runs automatically repair recognized original local paths and cached prediction identities before cache lookup, preserving PDF and prediction content. Conflicting layouts or malformed legacy metadata stop before inference. No separate migration command is required for normal Pro use. Release framework support before changing HF main.

Framework review: https://github.com/run-llama/ExtractBench/pull/101
Release checklist: https://github.com/run-llama/ExtractBench/issues/102

Leave unmerged. No HF main update, Hub publication, or model inference.

xujustinj changed pull request title from Clarify ExtractBench Pro challenge display names to Rename Pro splits, case IDs, document paths, and Harbor tasks
xujustinj changed pull request status to merged

Sign up or log in to comment