STAGE-Eval - Value Accuracy: leaderboard
STAGE-Eval (Ahn, Son and Yu, 2026) tests extracting structured JSON from documents: 851 held-out items each give a report written around a real spreadsheet, an extraction prompt and a JSON Schema, with gold values checked against the spreadsheet. Value accuracy in percent, the mean share of gold leaf values reproduced at the same JSON path, averaged over three runs with a 3,100-token cap; higher is better. The authors' fine-tunes trained on STAGE data are in-distribution.
Metric: Value accuracy (%; mean over items of the share of gold leaf values the prediction matches at the same JSON path; the 851-item held-out STAGE-Eval test set: an LLM-written report around a real spreadsheet, an extraction prompt and a JSON Schema per item; vLLM, temperature 0.6, at most 3,100 new tokens, JSON read after any reasoning; mean of three runs, the standard deviation printed run-on in the same cell not read). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 3B Instruct | 45.87 | #5150 |
| 2 | Qwen 3 4B (Non-reasoning) | 45.46 | #2281 (Qwen 3 4B) |
| 3 | Llama 3.2 3B Instruct | 36.05 | #7147 |
| 4 | Llama 3.2 1B Instruct | 19.54 | #9047 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=stage-eval-value-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-13.