Datasets:

---

llamaindex

/

ExtractBenchPro

Like 3

Follow

LlamaIndex 331

Adds 100 Harbor tasks generated from the existing public cases, using the shared Claude Code/Codex prompt from ExtractBench Appendix C.1, with the staged document filename standardized.

• Existing JSONL files, schemas, source documents, and dataset configuration are preserved.
• Agents see the document/schema and write output.json; grading runs separately through the existing evaluator at pinned commit 2f27f100730afc7c395602e26a2a87e07ea66b31.
• Agent timeout: 1,800 seconds per document for both ExtractBench and ExtractBenchPro. The verifier has a separate 1,800-second budget.
• Reward: Value F1. The prompt does not request citations, so this lane does not measure grounding.
• harbor-manifest.json records source revision 16f6672e06cf77491ec6d18c817d67864b90d845, scorer revision, original case IDs, and file hashes.
• The README adds Harbor usage/discovery tags, and .gitattributes adds LFS handling for task PDFs.

The verifier writes feedback.json with aligned reference/output cell paths, the outcome (matched, value_differs, reference_only, or prediction_only), and each cell's f1_credit. Records come from the same scoring pass, including reordered and nested tables. Matched cells receive 2 / (expected_cells + predicted_cells) credit; other cells receive zero. Credits sum to the existing document reward. Paths address normalized scoring inputs, including schema defaults and layout reconciliation. Reference values are not included. This explains the scored output; it does not assign marginal edit rewards or reasoning/token credit.

Validation completed:

• All 470 generated task configs were checked for a consistent 1,800-second agent limit. The original dataset configs and manifest have been corrected; Pro already used this limit.

• All 100 task configurations and file hashes validated. This update changes each task's scorer metadata, verifier Dockerfile, and verifier adapter; reference files, PDFs, schemas, instructions, and solutions are unchanged.

• Ordinary Hugging Face loading verified for all 6 splits; all source JSONL files read back unchanged from the new PR revision.

• All 18 saved reference/partial/empty comparisons preserve normal evaluator scores, and cell credits reconcile with the reward.

• Updated local Docker and clean-runner container smokes pass oracle=1 and nop=0 for both datasets, including feedback collection, without model calls.

• Tooling PR: 1,093 local tests pass; source CI passes on Python 3.12/3.13. Formatting, lint, type checks, package build, and red/green regression checks pass.

• The uploaded manifest, README, all source JSONL files, and a complete representative task were read back and verified against local hashes.

• Harbor's remote --repo path was also validated on the earlier task revision; Git LFS prerequisites are documented.

The dataset default branch is unchanged until this PR is merged.

Install Git LFS and run git lfs install before the first remote run. To preview one task from this PR:

harbor run --repo https://huggingface.co/datasets/llamaindex/ExtractBenchPro@f32fd1d7d5056ee12286afa70e4d564fc353102d -p tasks -i house-counting-rules--vendor-statement-reconciliation-01-b7dd86993c -a oracle -n 1