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: 4
Gradient accumulation: 4
Learning rate: 5e-05
Scheduler: cosine
Warmup ratio: 0.03
Save steps: 500
Selected checkpoint: checkpoint-500
Selection reason: best validation avg_wer_normalized_capped (0.557665, corpus WER 0.659176)

Final Test Metrics

Checkpoint Selection Evidence

checkpoint	wer_normalized	cer_normalized	avg_wer_normalized	avg_cer_normalized	eval_loss
checkpoint-500	0.5576649148066538	0.2691629617628126	0.5576649148066538	0.2691629617628126	0.4462866187095642
checkpoint-1000	0.5581921038683963	0.2746129558293713	0.5581921038683963	0.2746129558293713	0.47903984785079956
checkpoint-1107	0.5603334157402827	0.27484894993170134	0.5603334157402827	0.27484894993170134	0.47903984785079956

severity_speech_impairment	n_samples	n_speakers	mean_wer	mean_cer	median_wer	median_cer
Severe (frequent breakdowns)	315	3	0.5507	0.2787	0.5625	0.25
Moderate (requires effort to understand)	347	3	0.4128	0.1637	0.375	0.1146
Mild (easily understood with minimal effort)	365	3	0.3348	0.117	0.2857	0.0714

Result folder: results/checkpoint-500/
Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped metadata analyses.