Leaderboard submission: Consilience parsing pipeline (Gemini 3.6 Flash)
I would like to submit the Consilience parsing pipeline, powered by Gemini 3.6 Flash, for evaluation/listing on DP-Bench.
On the released 200-document corpus and official scorer, a single locked full-corpus run scored:
| Metric | Score |
|---|---|
| TEDS | 96.34 |
| TEDS-S | 98.41 |
| NID | 95.00 |
| Table-F1 | 93.81 |
All 200/200 documents completed successfully, with 15.25 seconds mean per-document latency. The currently listed best TEDS and TEDS-S values are 96.06 and 97.62, respectively.
The run used google/gemini-3.6-flash through OpenRouter, Consilience's OCR prompt, a DP-Bench semantic-serialization instruction that distinguishes visually rendered charts/indexes from actual tables, and a deterministic Markdown-to-DP-Bench schema adapter.
No ground-truth lookup at inference time, per-document model/output selection, manual correction, or post-hoc output repair was used. The run had 55 reference tables, 58 predicted tables, and 53 matches (5 false positives, 2 false negatives).
This is a parsing-pipeline submission; I am not claiming that I trained the underlying Gemini model. The prediction JSON, complete per-document scorer output, configuration, reproduction command, and SHA-256 hashes are ready. Please let me know whether you prefer these as discussion attachments or through a pull request updating the leaderboard.