llamaindex/ExtractBench-results

ExtractBench results

Per-document inference outputs of every pipeline on the ExtractBench 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 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>/short.jsonl   one line per document, see schema/record.v1.schema.json
results/<pipeline>/medium.jsonl
results/<pipeline>/long.jsonl
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 <split>/<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.

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

pip install -U huggingface_hub
hf download llamaindex/ExtractBench-results --repo-type dataset --local-dir eb-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                           # ground truth into ./data
python ../eb-results/tools/explode.py --benchmark extractbench ../eb-results/results/llamaextract_cost_effective output/llamaextract_cost_effective
uv run extract-bench run llamaextract_cost_effective --skip_inference

The run reads output/<pipeline>/<split>/<stem>.result.json, scores every document, and writes one report per split. Pass --group short to score one split.

Submit your pipeline

1. Register the pipeline in the ExtractBench package and run it: uv run extract-bench run <pipeline>.
2. Pack and check:

   python tools/pack.py --benchmark extractbench --pipeline <pipeline> output/<pipeline> results/<pipeline>
   python tools/validate.py --benchmark extractbench 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:

   hf upload llamaindex/ExtractBench-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 dataset card.