PlantMarkerBench - Maize Valid F1: leaderboard
Validity F1 for gene–cell-type marker evidence on PlantMarkerBench’s balanced 600-sentence Maize pilot split, from Table 3 of the v2 paper. Native fractions are converted to percent exactly. Fifteen open models run through Ollama with the default prompt; two OpenAI models use direct prompting, at temperature zero. The always-valid baseline scores about 66.7 F1. This is distinct from evidence-type macro-F1 and from best-prompt or other-species results.
Metric: Marker-evidence validity F1 (%). Source: arxiv.org. Saturation forecast: Around May 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 45.1 | #228 |
| 2 | GPT-5.4 | 40.9 | #83 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=plantmarkerbench-maize-valid-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.