Doc2DB-Bench - Relation Tables (Oracle Entities): leaderboard
Cell-level F1 of the relationship tables in Doc2DB-Bench (2026) when the model is given the gold entity tables, isolating how well it links entities across tables from a long document; percent, higher is better.
Metric: Relation-level cell F1 (%; relationship tables predicted from the document, the schema and the ground-truth entity tables (oracle entity setting); 203 synthesized long-document instances over 42 database schemas from BIRD and Spider in seven domain groups; identical prompts, greedy decoding at temperature 0; cells aligned by global maximum-weight tuple matching, a cell matching on exact numeric equality or at least 90 percent string similarity). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 72.33 | #123 |
| 2 | Gemini 2.5 Pro | 65.46 | #266 |
| 3 | Claude Opus 4.6 | 63.17 | #102 |
| 4 | Gemini 2.5 Flash | 59.5 | #453 |
| 5 | Qwen 3 Max | 55.96 | #403 |
| 6 | GPT-4o | 50.68 | #741 |
| 7 | DeepSeek V4 Flash | 48.93 | #258 |
| 8 | Qwen 2.5 72B Instruct | 31.98 | #1092 |
| 9 | Qwen 2.5 14B Instruct | 28.92 | #1732 |
| 10 | Llama 3.1 70B Instruct | 16.97 | #1546 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=doc2db-bench-relation-tables-oracle-entities · How It Works · Data refreshed daily, snapshot 2026-11-13.