ALBA - Syntax: leaderboard
Metric: Mean judge score (0-100) on ALBA's 100 expert-written pt-PT syntax questions (sentence constituents, clause types, active and passive voice and syntax-driven proofreading), judged by Gemini-2.5-Pro with an expert pt-PT rubric (precision, linguistic quality, completeness), three expert-rated few-shot examples and greedy decoding, 1-5 rating rescaled to 0-100; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 85.8 | #145 |
| 2 | GPT-5 | 78.3 | #91 |
| 3 | Qwen 3 8B | 70.3 | #667 |
| 4 | Gemma 3 12B (IT) | 58 | #655 |
| 5 | Apertus-8B-Instruct-2509 | 47.5 | #1048 |
| 6 | Ministral-8B-Instruct-2410 | 45.8 | #1080 |
| 7 | Gemma 2 9B (IT) | 44.8 | #774 |
| 8 | EuroLLM-9B-Instruct | 43.3 | #1043 |
| 9 | Qwen 2.5 7B Instruct | 42.8 | #846 |
| 10 | Llama 3.1 8B Instruct | 35 | #1018 |
| 11 | Mistral 7B Instruct | 32.5 | #1345 |
| 12 | salamandra-7B-instruct | 26.5 | #1390 |
| 13 | OLMo-2-1124-7B-Instruct | 24.5 | #1349 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=alba-syntax · How It Works · Data refreshed daily, snapshot 2026-10-11.