KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b

CDLI Qwen3-ASR Luganda Typical Speech Fine-tune

This repo contains the selected checkpoint checkpoint-21500 from the LG-QWEN3-ASR-TYPICAL-0P6B-T1 run.

Training Setup

• Base model: Qwen/Qwen3-ASR-1.7B
• Datasets: KasuleTrevor/lg_100hrs + dmusingu/yogera-dataset (Luganda split)
• Training language tag: Luganda
• Forced inference language: disabled
• Epochs: 2
• Batch size: 2
• Gradient accumulation: 4
• Learning rate: 2e-05
• Scheduler: cosine
• Warmup ratio: 0.03
• Save steps: 500
• Selected checkpoint: checkpoint-21500
• Selection reason: best validation normalized WER (0.224765, CER 0.048727)

Final Test Metrics

• Corpus WER (normalized): 0.277206
• Corpus CER (normalized): 0.069751
• Average utterance WER (normalized): 0.281266
• Average utterance CER (normalized): 0.070795

Checkpoint Selection Evidence

checkpoint	wer_normalized	cer_normalized	avg_wer_normalized	avg_cer_normalized	eval_loss
checkpoint-21500	0.2247654288357341	0.04872670420957655	0.24759332060982953	0.050578496844777644	0.1309615969657898
checkpoint-21000	0.2248447204968944	0.04906116797859432	0.24706433444541367	0.050849497756757775	0.1293729841709137
checkpoint-22000	0.22740848420774415	0.05144519461124421	0.24720479928345496	0.05044484913280002	0.12995050847530365
checkpoint-23000	0.22777851195982557	0.05318362838741794	0.24681062129286088	0.05057407977163512	0.1293202042579651
checkpoint-22500	0.22862428967886878	0.051946890264770854	0.24762721848805866	0.05091758169808502	0.1302434355020523
checkpoint-20500	0.2296815118276728	0.05375532808562272	0.2472921023622922	0.050774016709905556	0.1301114857196808

Source Dataset Breakdown

source_dataset	n_samples	mean_wer	mean_cer	median_wer	median_cer
lg100	2828	0.2813	0.0708	0.2222	0.0395

Artifacts

• Result folder: results/checkpoint-21500/
• Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped analyses.