Modelpublic

KasuleTrevor/cdli-qwen3-asr-lg-typical-textpre-0p6b-v2-lr1e4

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes11downloads

Model Card

CDLI Qwen3-ASR Luganda Typical Speech Fine-tune

This repo contains the selected checkpoint checkpoint-24500 from the LG-QWEN3-ASR-TYPICAL-TEXTPRE-0P6B-V2-LR1E4 run.

Training Setup

—Base model: KasuleTrevor/cdli-qwen3-asr-lg-textpre-0p6b-v2
—Datasets: KasuleTrevor/lg_100hrs + dmusingu/yogera-dataset (Luganda split)
—Training language tag: Luganda
—Forced inference language: disabled
—Epochs: 6
—Batch size: 8
—Gradient accumulation: 2
—Learning rate: 0.0001
—Scheduler: cosine
—Warmup ratio: 0.03
—Save steps: 500
—Selected checkpoint: checkpoint-24500
—Selection reason: best validation normalized WER (0.190353, CER 0.047268)

Final Test Metrics

—Corpus WER (normalized): 0.249439
—Corpus CER (normalized): 0.077364
—Average utterance WER (normalized): 0.250726
—Average utterance CER (normalized): 0.076617

Checkpoint Selection Evidence

checkpoint	wer_normalized	cer_normalized	avg_wer_normalized	avg_cer_normalized	eval_loss
checkpoint-24500	0.19035284789216333	0.047268286612115365	0.21259655916521408	0.04884738253040073	0.38649630546569824
checkpoint-25000	0.19096075062772566	0.04713216763635232	0.21404931274441566	0.04892694348420396	0.38632816076278687
checkpoint-25314	0.19162151447072817	0.04725273015774245	0.2147274298980036	0.04904672960672542	0.38632816076278687
checkpoint-24000	0.19180652834676887	0.04760275038113313	0.21440237743817556	0.04930352224013652	0.38645777106285095

Source Dataset Breakdown

source_dataset	n_samples	mean_wer	mean_cer	median_wer	median_cer
lg100	2828	0.2507	0.0766	0.2	0.0338

Artifacts

—Result folder: results/checkpoint-24500/
—Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped analyses.