DagbaniGPT — v1-kaggle
Small decoder-only transformer language model for Dagbani (Dagbanli, ISO 639-3
dag), a Mabia (Gur) language of northern Ghana. Part of a ground-up Dagbani NLP
research program: corpus, tokenizer, benchmarks, and models built from first
principles. Source: https://github.com/pious2847/Dagbani-NLP_V2
Metrics (held-out, document-level split, per-line protocol)
| Metric | Value |
|---|---|
| Bits/token | 5.3492 |
| Perplexity | 40.76 |
| Bits/char | 1.5015 |
| Parameters | 49.9M |
| Tokenizer | dag_unigram_16k (SentencePiece unigram, byte fallback) |
Reference baselines on the same split: KN token 3-gram PPL 67.94; KN char 5-gram 2.099 bits/char.
Training data
Public Dagbani text (Wikipedia CC BY-SA, GhanaNLP community corpora, Naden dictionary) plus research-use-only text that is not redistributed here. This repo contains model weights, tokenizer, and metrics only — no corpus data.
Usage
The checkpoint is a plain PyTorch state dict + config; see src/lm/train_lm.py
in the source repo for the DagbaniGPT class definition and loading example.