Dev trains 46M-param native Luganda model Pearl Chat, still collapses after 2 turns
weskambale · x · 2026-11-13
Developer weskambale reports training Pearl Chat, a new 46M-parameter natively Luganda model. Pretraining results show promising progress on syntax, spelling, and sentence structure, but multi-turn conversation remains weak: the model currently collapses after just 2 conversational turns.
With a limited SFT dataset, the team has launched a new data acquisition pipeline to collect more Luganda prose and conversations, and plans a next experiment with a 150M-parameter model. A screenshot shows the model answering Uganda geography questions in Luganda.
Related event: Developer Trains 46M-Parameter Luganda Model That Collapses After Two Turns(2 posts)→
More from Models
- Cloudflare hosts Thinking Machines' inkling 2 small on Workers AI at day-0 speed — michellechen · 2026-11-14
- Masked decoding trick tames text rivers in Gemma 3 12B, with a width-aware LoRA and coherence RL — cephaloform · 2026-11-14
- 250,000-human AI mafia game finds OpenAI models lie most, Anthropic snitches — Scobleizer · 2026-11-14
- Developer downgrades Claude Ultra to Pro and moves all work to Codex over 5-hour rate limits — umeshonai · 2026-11-14
- Pangram details per-domain false positive rates (1 in 10,000), enabling profitable AI detection — soldni · 2026-11-14
- Blogger predicts GPT-6.2 launches before Nov 19, floats three reasons for the delay — kimmonismus · 2026-11-14