Singlish Llama
Llama models fine-tuned for Singlish, the Sinhala and English mix Sri Lankans actually type.
The problem
General models stumble on code-mixed Sinhala and English, the way most people in Sri Lanka write online.
The approach
Curated code-mixed corpora and instruction sets, then continued pre-training and LoRA fine-tuning of Llama 3.1 8B and Llama 3.2 3B with Unsloth.
What stands out
- Led the sinGPT team to the final eight of 30 teams at the Commercial Bank international hackathon.
- Published five models and eight datasets on Hugging Face, the largest corpus with 10–100 million rows.
Measured
-
8 of 30
The sinGPT team was one of 8 finalists from 30 teams at the Commercial Bank international hackathon.
A Sinhala-English model built by students held its own in an international field.
Aug 2024