Add Harbor tasks using the published ExtractBench coding-agent prompt
Adds 370 Harbor tasks from the existing public cases, using the published coding-agent prompt. Original JSONL files, schemas, source documents, and dataset configuration are preserved.
Agents receive the PDF/schema, write output.json, and have 1,800 seconds per document. Agent-phase networking permits only explicitly allowed model API hosts. The separate offline verifier uses the pinned existing evaluator; Value F1 is the sole reward. feedback.json attributes it to aligned cells in normalized scoring inputs.
The manifest records source/scorer revisions and task hashes. All 370 tasks were checked against their source rows; uploaded manifests, sample configs, README, and every source JSONL were read back and verified. Tooling and validation.
With Harbor 0.24.0+, Docker, and Git LFS installed (git lfs install), preview one task without model calls:
harbor run --repo https://huggingface.co/datasets/llamaindex/ExtractBench@3d8569400fbb04bdeffde38e12faab14bf4ad428 -p tasks -a oracle -l 1 -n 1
For extraction, use -a codex -m MODEL --allow-agent-host MODEL_API_HOST instead of -a oracle. Allow only required model API hostnames. This prompt does not measure grounding.