---
license: apache-2.0
viewer: false
pretty_name: ExtractBench Pro results
tags:
- document-extraction
- benchmark
- evaluation
- json-schema
---

# ExtractBench Pro results

Per-document inference outputs of every pipeline on the [ExtractBench Pro](https://huggingface.co/datasets/llamaindex/ExtractBenchPro) leaderboard, so anyone can re-run the evaluation locally without calling any provider, inspect what a pipeline returned for a given document, and re-score when the ground truth changes.

Scores come from the [ExtractBench](https://github.com/run-llama/ExtractBench) package. Pipeline results were produced by the same package or by the equivalent provider API calls; only the normalized output and the cost, usage and timing counters are published.

## Layout

```
pipelines.jsonl                        one line per pipeline: name, display name, category, documents, data revision
results/<pipeline>/<challenge>.jsonl   one line per document per challenge, see schema/record.v1.schema.json
                                       (document_conventions, house_vocabulary, house_counting_rules,
                                        drawn_relations, charts_inside_documents, long_chart_sheets)
data_manifest.json                     document ids of the data revision these results were scored on
tools/                                 explode.py, pack.py, validate.py (Python 3.9+, standard library)
```

Each line is a minimal `InferenceResult` as the evaluator loads it: `request.example_id` is `<challenge>/<document stem>`, `output` holds `extracted_data` and `field_citations` (the citations drive the word- and page-grounding scores), and `raw_output` holds cost, page and token counters. Records carry the real wall-clock latency of each document; the leaderboard's 30-minute view is a cutoff over `latency_in_ms`, the no-time-limit columns use every record. Documents a pipeline failed on have no record and are scored as missing.

## Two tiers: eval records and full results

- `results/<pipeline>/<category>.jsonl` is the small tier: only the fields the evaluator needs. Download this to re-score a pipeline quickly.
- `raw/<pipeline>/<category>.jsonl` is the full tier: one line per document with everything a user of that pipeline receives, meaning the complete provider response (`raw_output`), the normalized output with all citation metadata, the parse job spawned by LlamaExtract (`stages`), timing and cost. Fields that exist only because of our harness (internal paths, run ids, job ids, signed download links) are removed.
- `traces/<pipeline>/<category>/<stem>.*` holds the complete transcripts of coding-agent pipelines: the prompt, every tool call and tool output, thinking, token usage and cost (`*.trace.jsonl`, `*.attempt-N.jsonl`), the agent's final message (`*.last_message.txt`), its structured output (`*.output.json`) and stderr (`*.stderr.log`). Only our runner paths and identity fields are rewritten or removed.

Download one tier at a time with `--include "results/<pipeline>/*"`, `--include "raw/<pipeline>/*"` or `--include "traces/<pipeline>/*"`.

## Replay a pipeline locally

```bash
pip install -U huggingface_hub
hf download llamaindex/ExtractBenchPro-results --repo-type dataset --local-dir pro-results \
  --include "results/llamaextract_cost_effective/*" "tools/*" "data_manifest.json" "pipelines.jsonl"

git clone https://github.com/run-llama/ExtractBench && cd ExtractBench && uv sync
uv run extract-bench download --dataset pro             # ground truth into ./data-pro
python ../pro-results/tools/explode.py --benchmark extractbenchpro ../pro-results/results/llamaextract_cost_effective output-pro/llamaextract_cost_effective
uv run extract-bench run llamaextract_cost_effective --dataset pro --skip_inference
```

The run reads `output-pro/<pipeline>/<challenge>/<stem>.result.json` and writes one report per challenge. Pass `--group document_conventions` to score one challenge.

## Submit your pipeline

1. Register the pipeline in the ExtractBench package and run it: `uv run extract-bench run <pipeline> --dataset pro`.
2. Pack and check:
   ```bash
   python tools/pack.py --benchmark extractbenchpro --pipeline <pipeline> output-pro/<pipeline> results/<pipeline>
   python tools/validate.py --benchmark extractbenchpro results/<pipeline>
   ```
   `validate.py` fails on anything that is not a public document id, on local paths or hostnames, on credentials, and on non-canonical JSON. Identical model output always packs to identical files; only timestamps, latency and cost differ between runs.
3. Open a pull request on this dataset:
   ```bash
   hf upload llamaindex/ExtractBenchPro-results results/<pipeline> results/<pipeline> --repo-type dataset --create-pr
   ```
   and link it from your leaderboard pull request on GitHub. The leaderboard row is re-scored from the uploaded results before merge.

## Citation

See the [ExtractBench Pro dataset card](https://huggingface.co/datasets/llamaindex/ExtractBenchPro).