KasuleTrevor/cdli-qwen3-asr-lg-atypical-stage3-1p7b-base-b8

CDLI Qwen3-ASR Luganda Atypical Speech Fine-tune

This repo contains the selected checkpoint checkpoint-1000 from the LG-QWEN3-ASR-ATYPICAL-STAGE3-1P7B-BASE-B8 run.

Training Setup

• —Base model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
• —Dataset: cdli/ugandanlugandanonstandardspeechv1.0
• —Split files: train_cleaned.tsv, validation_cleaned.tsv, test_cleaned.tsv
• —Training language tag: Luganda
• —Forced inference language: disabled
• —Epochs: 10
• —Batch size: 4
• —Gradient accumulation: 4
• —Learning rate: 0.0001
• —Scheduler: cosine
• —Warmup ratio: 0.03
• —Save steps: 500
• —Selected checkpoint: checkpoint-1000
• —Selection reason: fallback best eval_loss (0.512122)

Final Test Metrics

• —Primary WER: avg utterance WER normalized, capped at 1.0 = 0.497477
• —Primary CER: avg utterance CER normalized, capped at 1.0 = 0.230633
• —Corpus WER (normalized, uncapped): 0.633333
• —Corpus CER (normalized, uncapped): 0.322517

Checkpoint Selection Evidence

Severity Breakdown

Artifacts

• —Result folder: results/checkpoint-1000/
• —Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped metadata analyses.