Base model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
Split files: train_cleaned.tsv, validation_cleaned.tsv, test_cleaned.tsv
Training language tag: Luganda
Forced inference language: disabled
Epochs: 3
Batch size: 2
Gradient accumulation: 4
Learning rate: 2e-05
Scheduler: cosine
Warmup ratio: 0.03
Save steps: 500
Selected checkpoint: checkpoint-1500
Selection reason: best validation avg_wer_normalized_capped (0.593081, corpus WER 0.714471)

Final Test Metrics

Checkpoint Selection Evidence

checkpoint	wer_normalized	cer_normalized	avg_wer_normalized	avg_cer_normalized	eval_loss
checkpoint-1500	0.5930810256197367	0.3011593535760467	0.5930810256197367	0.3011593535760467	0.4881881773471832
checkpoint-2000	0.5946540381673293	0.30105220481571093	0.5946540381673293	0.30105220481571093	0.48891597986221313
checkpoint-2211	0.5961148991272694	0.30305782850209245	0.5961148991272694	0.30305782850209245	0.48891597986221313
checkpoint-1000	0.5985111617145635	0.3072198319719475	0.5985111617145635	0.3072198319719475	0.49404463171958923
checkpoint-500	0.6380178007084313	0.3405655194402582	0.6380178007084313	0.3405655194402582	0.5250704288482666

severity_speech_impairment	n_samples	n_speakers	mean_wer	mean_cer	median_wer	median_cer
Severe (frequent breakdowns)	315	3	0.6088	0.3306	0.625	0.28
Moderate (requires effort to understand)	347	3	0.4166	0.1704	0.4	0.132
Mild (easily understood with minimal effort)	365	3	0.369	0.1391	0.3333	0.0847

Result folder: results/checkpoint-1500/
Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped metadata analyses.