llamaindex/ParseBench-2-results

ParseBench-2 results

Per-document inference outputs of every pipeline on the ParseBench-2 leaderboard, so anyone can re-run the evaluation locally without calling any provider, inspect what a pipeline returned for a given page, and re-score when the ground truth or the scorers change.

Scores come from the ParseBench package. 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>/<subset>.jsonl     one line per document, for each of the 13 subsets (text, headings, strikeout,
                                      underline, tables, complex_table_headers, financial_tables,
                                      insurance_filing_tables, dense_tables, charts, layout, forms, fillable_forms);
                                      see schema/record.v1.schema.json
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 <subset>/<document name>, the dataset's own id. output is the normalized parse output (markdown, pages, layout) and raw_output holds cost, page and token counters. Layout boxes (layout_pages[].items and granular_layers) are kept only in layout.jsonl, the one subset that scores them; elsewhere the scorers read only each page's number and header/footer Markdown. For layout records of providers whose layout adapter reads the provider payload, raw_output also carries that payload with harness fields removed.

Replay a pipeline locally

pip install -U huggingface_hub
hf download llamaindex/ParseBench-2-results --repo-type dataset --local-dir pb2-results \
  --include "results/azure_di_layout/*" "tools/*" "data_manifest.json" "pipelines.jsonl"

git clone https://github.com/run-llama/ParseBench && cd ParseBench && uv sync
python ../pb2-results/tools/explode.py --benchmark parsebench-2 ../pb2-results/results/azure_di_layout output/parsebench-2
uv run parse-bench run azure_di_layout --dataset parsebench-2 --skip_inference

explode.py writes output/parsebench-2/<subset>/<pipeline>/<name>.result.json, the layout parse-bench run --dataset parsebench-2 reads. The run downloads the dataset into ./data_parsebench-2 if needed, scores every subset and writes the six dimension scores and the overall score to output/parsebench-2/_score/<pipeline>/_next_score.json. Pass --group charts,tables to score some subsets.

Submit your pipeline

1. Register the pipeline in the ParseBench package and run it: uv run parse-bench run <pipeline> --dataset parsebench-2.
2. Pack and check:

   python tools/pack.py --benchmark parsebench-2 --pipeline <pipeline> output/parsebench-2 results/<pipeline>
   python tools/validate.py --benchmark parsebench-2 results/<pipeline>
   

   validate.py fails on anything that is not a public document id, on local paths or hostnames, on credentials, on layout boxes outside layout.jsonl, 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:

   hf upload llamaindex/ParseBench-2-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 ParseBench-2 dataset card.