parsebench-results / README.md
ianrutneratyou's picture
Update README.md
769ebc2 verified
|
Raw History Blame Contribute Delete
5.48 kB
---
license: apache-2.0
viewer: false
pretty_name: ParseBench results — URL extract pipelines
tags:
- document-parsing
- benchmark
- evaluation
- ocr
---
# ParseBench results — URL extract pipelines
Per-document outputs of five URL-extract pipelines on the [ParseBench](https://huggingface.co/datasets/llamaindex/ParseBench) benchmark. Use these outputs to re-run the evaluation locally without calling any provider, inspect what a pipeline returned for a given page, or re-score when the ground truth or the scorers change.
The outputs were produced 2026-11-04 to 2026-11-09 with a You.com working copy of the [ParseBench](https://github.com/run-llama/ParseBench) harness (`parse_bench` 1.0.7, base commit `6c5a2f4`).
## Layout
```
pipelines.jsonl one line per pipeline: name, display name, category, documents, data revision
raw/<pipeline>/<subset>.jsonl full tier: one line per document — provider payload (sanitized) + normalized output
results/<pipeline>/<subset>.jsonl eval tier: one line per document, see schema/record.v1.schema.json
results/<pipeline>/failed.json documents that produced no output: {"<subset>": [{"id": ..., "error": ...}]}
data_manifest.json document ids of the data revision these outputs were produced against
schema/record.v1.schema.json record schema v1, verbatim copy from upstream
tools/ explode.py, pack.py, validate.py, verbatim copy from upstream
```
Each JSONL line is a minimal `InferenceResult` as the public evaluator loads it: `request.example_id` is `<subset>/<document stem>`, `output` is the normalized parse output (markdown, pages, layout), and `raw_output` holds page and usage counters. Provider cost fields are intentionally not published. Subsets: `chart`, `table`, `text` (`text` is scored as both text_content and text_formatting; the `layout` subset was not run).
## Two tiers: eval records and full results
- `results/<pipeline>/<subset>.jsonl` is the small tier: only the fields the evaluator needs. Download this to re-score a pipeline quickly.
- `raw/<pipeline>/<subset>.jsonl` is the full tier: one line per document with everything a caller of that pipeline receives — the complete provider response plus the normalized output. Harness-only fields (internal paths, job ids, signed download links) and provider cost fields are removed.
Download one tier at a time with `--include "results/<pipeline>/*"` or `--include "raw/<pipeline>/*"`.
## Pipelines
| pipeline | display name | provider |
|---|---|---|
| `youcom_extract` | You.com Extract | youcom |
| `youcom_extract_lite` | You.com Extract (parse_effort: lite) | youcom |
| `tavily_extract` | Tavily Extract | tavily |
| `exa_contents` | Exa Contents | exa |
| `parallel_extract` | Parallel Extract | parallel |
Documents that failed provider-side produce no JSONL record; they are listed per subset in `results/<pipeline>/failed.json`.
## Re-score
From a checkout of [youdotcom-oss/youdotcom-parsebench-eval](https://github.com/youdotcom-oss/youdotcom-parsebench-eval) or of [run-llama/ParseBench](https://github.com/run-llama/ParseBench) `main`, with `HF_TOKEN` set while this repo is private. These pipelines are not `leaderboard.csv` rows, so `scripts/leaderboard/rescore.py` gets a results map that points at this repo (`results_repo` is a top-level map key; there is no CLI flag for it):
```bash
cat > /tmp/youdotcom-results.yaml <<'EOF'
results_repo: youdotcom/parsebench-results
data_repo: llamaindex/ParseBench
data_revision: 2805a1d940f95a203e0ae4b88be9934f7765b3fc
scored_with: null
rows:
- {provider: You.com Extract, status: verified, pipeline: youcom_extract, results_revision: c34b1b96dbab99a92c28f2beaa256f56f035dcdf}
- {provider: You.com Extract Lite, status: verified, pipeline: youcom_extract_lite, results_revision: c34b1b96dbab99a92c28f2beaa256f56f035dcdf}
- {provider: Tavily Extract, status: verified, pipeline: tavily_extract, results_revision: c34b1b96dbab99a92c28f2beaa256f56f035dcdf}
- {provider: Exa Contents, status: verified, pipeline: exa_contents, results_revision: c34b1b96dbab99a92c28f2beaa256f56f035dcdf}
- {provider: Parallel Extract, status: verified, pipeline: parallel_extract, results_revision: c34b1b96dbab99a92c28f2beaa256f56f035dcdf}
EOF
uv run python scripts/leaderboard/rescore.py run --map /tmp/youdotcom-results.yaml \
--pipelines youcom_extract_lite --out rescore.json
```
Pass any comma-separated subset of the five pipelines to `--pipelines`. The first run downloads the ground truth into `.cache/leaderboard-gt/`; after that, one pipeline takes roughly 10–15 minutes. Scores land in `rescore.json` under `results.<pipeline>`: `scores` (Tables, Charts, Content_Faithfulness, Semantic_Formatting, Visual_Grounding), `overall`, and `details.<column>` (published vs. ground-truth page counts and any error). `layout` was not run, so `Visual_Grounding` is 0 and is included in `overall`. Documents listed in `failed.json` have no record and score 0. `apply.py` is not needed, since it only updates `leaderboard.csv` rows.
`tools/` is a verbatim copy of the same directory from [llamaindex/ParseBench-results](https://huggingface.co/datasets/llamaindex/ParseBench-results) (Apache-2.0). The re-score flow downloads `tools/explode.py` from the results repo at the pinned revision, so it must live in this repo.
## Citation
See the [ParseBench dataset card](https://huggingface.co/datasets/llamaindex/ParseBench).