HELM SeaHELM - TyDiQA: leaderboard
HELM SeaHELM Southeast-Asian-languages evaluation: typologically-diverse multilingual question answering.
Metric: SQuAD macro-averaged F1 score. Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash | 81.51 | #342 |
| 2 | Gemini 2.0 Flash Lite | 65.01 | #450 |
| 3 | Llama 4 Maverick Instruct FP8 | 64.9 | #414 |
| 4 | GPT-4o (2024-11-20) | 63.68 | #381 |
| 5 | Gemma 2 9B (IT) | 62.17 | #788 |
| 6 | Llama 3.2 90B Vision Instruct | 60.07 | #622 |
| 7 | Llama 3.3 70B Instruct | 57.34 | #534 |
| 8 | Gemma 2 27B (IT) | 56.32 | #645 |
| 9 | Llama 4 Scout Instruct | 55.91 | #539 |
| 10 | GPT-4o Mini (2024-07-18) | 52.46 | #558 |
| 11 | Llama 3.1 405B Instruct | 52.2 | #457 |
| 12 | Llama 3.1 70B Instruct | 51.04 | #561 |
| 13 | Qwen 2 72B Instruct | 50.52 | #576 |
| 14 | DeepSeek R1 Distill Llama 8B | 50.19 | #1291 |
| 15 | Llama 3.1 8B CPT SEA LIONv3 Instruct | 48.87 | #810 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-seahelm-tydiqa · How It Works · Data refreshed daily, snapshot 2026-10-20.