AI2 Releases Mangosteen: A High-Quality Thai Pretraining Corpus
allen_ai · x · 2026-08-27
AI2 released Mangosteen, a 47B-token Thai pretraining corpus. Models trained on it matched or beat Thai LLMs trained on larger web datasets after filtering duplicates and low-quality text. These gains held with larger models. AI2 emphasizes making tools like Dolma open to help researchers adapt to specific languages and communities.
Related event: Ai2 Helps Build 47B-Token Thai Corpus Mangosteen(2 posts)→
More from Models
- MiniMax H3 Max tops video leaderboards via fal's post-training — ArtificialAnlys · 2026-08-27
- GLM-5.3 Flash Review: GPT-5.6 Level Performance at Ultra-Low Cost — zainhas · 2026-08-27
- Users Disappointed with Qwen3.8 Flash Next Release — NigaTroubles · 2026-08-27
- Qwen 3.8 and GLM 5.3 Flash open models released, available for Dell on-prem deployment — _akhaliq · 2026-08-27
- Fal secures new funding as H3 Max model achieves fastest speed and best quality — isidentical · 2026-08-27
- Local model choice in 2026: Qwopus 9B vs Qwen 27B — Ammargok · 2026-08-27