FinixDocBench_Eval_for_Markdown/README.md

FinixDocBench Markdown Evaluation

This folder contains a lightweight evaluator for FinixDocBench Markdown parsing outputs. It compares a directory of ground-truth .md files with a directory of predicted .md files whose file names match exactly.

The evaluator is intended for the public Markdown task. It does not evaluate structured JSON layout annotations.

Metrics

The evaluator reports:

Metric	Direction	Meaning
text_block_Edit_dist	Lower is better	Normalized edit distance over matched text blocks.
reading_order_Edit_dist	Lower is better	Normalized edit distance over serialized reading-order sequences.
table_TEDS	Higher is better	TEDS table-structure similarity, scaled to 0-100.
overall	Higher is better	Composite score on a 0-100 scale.

The overall score is:

overall = (
    (1 - text_block_Edit_dist) * 100
    + (1 - reading_order_Edit_dist) * 100
    + table_TEDS
) / 3

Formula parsing is not evaluated separately.

Installation

Python 3.9+ is recommended.

cd FinixDocBench_Eval_for_Markdown
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Quick Check

Run the bundled minimal example first:

python run_eval.py \
  --gt_dir examples/gt \
  --pred_dir examples/pred \
  --output_json outputs/example_result.json

Evaluate a FinixDocBench Track

Prepare a prediction directory with one .md file per evaluated page. Prediction file names must match the ground-truth Markdown file names.

Example for FinixPhoto:

python run_eval.py \
  --gt_dir ../track2_finixphoto_300/mds \
  --pred_dir /path/to/predicted_mds \
  --output_json outputs/finixphoto_result.json

Example for FinixHuge-Table:

python run_eval.py \
  --gt_dir ../track3_finixhuge_100_table/mds \
  --pred_dir /path/to/predicted_mds \
  --output_json outputs/finixhuge_table_result.json

Output Format

The output JSON has the following structure:

{
  "success": true,
  "metrics": {
    "text_block_Edit_dist": 0.0123,
    "reading_order_Edit_dist": 0.0,
    "table_TEDS": 98.7,
    "overall": 99.15,
    "num_samples": 2,
    "score": 99.15
  },
  "inputs": {
    "gt_files": 2,
    "pred_files": 2,
    "missing_predictions": 0,
    "unexpected_predictions": 0
  }
}

score is identical to overall and is included for leaderboard or automation systems that expect a generic score field.

File Name Validation

By default, the evaluator fails if the .md file names in gt_dir and pred_dir do not match exactly. This avoids accidentally skipping pages.

If you want to allow missing predictions and score missing files as empty outputs, pass:

python run_eval.py \
  --gt_dir /path/to/gt_mds \
  --pred_dir /path/to/pred_mds \
  --allow_name_mismatch

FinixHuge Reporting

For FinixHuge-Long and FinixHuge-Table, also report a success rate outside this script:

success_rate = valid_non_empty_predictions / total_pages

A prediction should be counted as successful only if it is a syntactically valid, non-empty page-level Markdown result without runtime failure, severe truncation, or format errors that prevent downstream evaluation.

Large Table Safeguards

TEDS can be slow on extremely large tables. To keep evaluation practical, this implementation assigns a table TEDS score of 0 when a ground-truth or predicted table exceeds 50000 <td> cells.

The matching stage also keeps broad safety thresholds:

MAX_PRED_ITEMS = 50000
RATIO_THRESHOLD = 100
MAX_TOTAL_LENGTH = 10000000
MAX_SINGLE_ITEM_LENGTH = 10000000

Notes

• Only .md files are evaluated.
• Images, JSON annotations, and other files are ignored by this evaluator.
• Markdown tables are converted to HTML before table evaluation.
• One page image should correspond to one Markdown file.
• File names are the matching keys; the evaluator does not read images.