Rename split files, case-ID prefixes, PDF paths, category tags and card links. Preserve all 100 PDFs, schemas, reference values and scoring rules. Harbor task names remain based on document identifiers.
Remove challenge names from Pro Harbor task identities
Derive task slugs and hashes from document identifiers, retaining source split and case ID in metadata. All 100 execution payloads remain byte-identical. No legacy task-name aliases.
Regenerate Harbor tasks with canonical Pro case IDs
Replace the 78 renamed task directories with canonical IDs. Regenerate all task identity metadata and the manifest against the reviewed data commit. All 100 task execution payloads, schemas, PDFs, oracle answers, verifier code, and scorer pin match the original release; only identity metadata and challenge tags change. No legacy task aliases are included.
Use canonical Pro split names, case IDs, and document paths
Rename five split files and 78 case/PDF paths. Preserve all PDF bytes and schema, expected-output, scoring-rule, and repeated-structure fields. Add the complete old-to-new case map with unchanged PDF SHA-256 hashes. Harbor overlay follows in the same unmerged PR.
Use task- and format-based challenge labels consistently across the dataset card: Charts, Relational diagrams, Document conventions, Custom terminology, Long dense pages, and Counting.
The labels identify visual structure, document-specific definitions, aggregation, or page geometry directly. They remove the ambiguous qualifier “house” and shorten redundant wording. Existing split identifiers remain stable so published commands, case IDs, file paths, and Harbor tasks retain their identity.
Only README prose changes. The YAML dataset configuration, all JSONL/PDF files, task packages, manifest, split membership, and scoring are unchanged. No benchmark rerun is needed.
Validation:
- [x] Compared the YAML front matter byte-for-byte with the current public card.
- [x] Reviewed the diff: only five display-name replacements and a compatibility note.
- [x] Upload contains README.md only and is guarded by the source commit.
Leave this PR unmerged pending coordinated review.
18 days ago
Restore RL Environment discovery for ExtractBench Pro (#6)
- Add Harbor tasks using the published ExtractBench coding-agent prompt (8b5daf42e2adb5e8ef6f13d0bc4366732cd7970f)
- Document Git LFS prerequisite for Harbor (092b835ca9918faf75c8c97e041326a10f2686ce)
- Add aligned cell feedback for the existing Value F1 reward (f32fd1d7d5056ee12286afa70e4d564fc353102d)
- Enforce agent-phase model-host allowlists and offline verification (f5eb7df6f3f9ceb635e6cec805e10ee28c68bb03)
Subsets, folders, row ids and challenge tags now follow the six challenges: document conventions, house vocabulary, house counting rules, drawn relations, charts inside documents and long chart sheets. Documents, schemas, expected outputs and field rules are unchanged.
