KasuleTrevor/cdli-qwen3-asr-lg-atypical-stage3-1p7b-typical-base-const1e4
Qwen3-ASR Luganda Atypical Stage 3 1.7B
This repository contains a Qwen3-ASR checkpoint fine-tuned for Luganda atypical speech from the CDLI Luganda non-standard speech dataset.
Training Setup
- Base model:
KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune - Dataset:
cdli/ugandan_luganda_nonstandard_speech_v1.0 - Language: Luganda (
lg) - Training source notebook:
qwen3-asr-luganda-atypical-cleaned-finetune.ipynb - Learning rate:
1e-4 - Scheduler:
constant_with_warmup - Epochs:
3 - Batch size:
2 - Gradient accumulation:
4 - Save steps:
250 - SpecAugment: not used
Checkpoint Selection
Notebook outputs show that validation checkpoint comparison selected checkpoint-2000 by normalized corpus WER.
A later exploratory notebook cell evaluated checkpoint-1250 after the upload. That post-hoc test result should not be interpreted as the uploaded best checkpoint.
Notebook-Recovered Test Artifact
The notebook contains test metrics for the later post-hoc checkpoint-1250 evaluation:
Per-Speaker Breakdown
These per-speaker aggregates come from the notebook-displayed grouped analysis for the recovered post-hoc checkpoint-1250 test artifact.
<!-- markdownlint-disable MD013 -->
<!-- markdownlint-enable MD013 -->
Evaluation Notes
The model card is reconstructed from executed notebook outputs and extracted local summaries. Full per-utterance prediction files were referenced in notebook outputs as remote /workspace/... artifact paths, but were not embedded in the notebook JSON available locally.
