FilBench - Generation: leaderboard
From the FilBench leaderboard (Miranda et al., 2025), which evaluates LLMs on Filipino, Tagalog and Cebuano with a lighteval fork. Generation combines the NTREX-128, Tatoeba (Cebuano and Tagalog) and TICO-19 translation tasks, scored with ROUGE-L and weighted by instance count; 0-100, higher is better. Runs missing one of these tasks are omitted.
Metric: Category Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 40 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 54.65 | #343 |
| 2 | GPT-4o Mini | 47.88 | #598 |
| 3 | Llama 4 Maverick Instruct FP8 | 33.67 | #424 |
| 4 | DeepSeek V3 | 27.53 | #329 |
| 5 | Llama 3.1 70B Instruct | 24.03 | #567 |
| 6 | Llama-SEA-LION-v3-70B-IT | 23.95 | #407 |
| 7 | Llama 4 Scout Instruct | 19.75 | #547 |
| 8 | gemma2-9B-cpt-sea-lionv3-instruct | 17.1 | #717 |
| 9 | Sailor2-20B-Chat | 15.95 | #589 |
| 10 | Gemma 2 27B (IT) | 15.38 | #652 |
| 11 | Qwen 2.5 72B Instruct | 14.98 | #469 |
| 12 | Sailor2-8B-Chat | 14.29 | #783 |
| 13 | Qwen 2 72B Instruct | 12.11 | #579 |
| 14 | Gemma 2 9B (IT) | 11.51 | #795 |
| 15 | Mixtral 8x22B Instruct | 10.7 | #738 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=filbench-generation · How It Works · Data refreshed daily, snapshot 2026-11-05.