serendepify-gsl-asr-ak-waxal-gnlp-whisper-small-balanced-fullft-v0.2
This is the second Ghanaian Speech Lab ASR artifact: a balanced Waxal+GhanaNLP full fine-tuning pass intended to improve WER beyond the Round 2 checkpoint.
It is a real training run, not a smoke-test checkpoint, but it is still a review candidate until the held-out evaluation and failure taxonomy are complete.
License is set conservatively because this pass includes Waxal-derived data.
- Training rows: 4000
- Dev rows: 256
- Max steps: 1200
- Method: full fine-tuning from
teckedd/whisper-small-waxal-round2-specaug-v1 - Dataset mix: balanced Waxal + GhanaNLP from the v0.1 sanitized manifest
Baseline metrics:
{
"baseline_loss": 5.176477432250977,
"baseline_model_preparation_time": 0.0045,
"baseline_wer": 0.4432036965729688,
"baseline_runtime": 126.0875,
"baseline_samples_per_second": 2.03,
"baseline_steps_per_second": 0.508
}
Final metrics:
{
"final_loss": 0.9554044008255005,
"final_model_preparation_time": 0.0045,
"final_wer": 0.5246438197920678,
"final_runtime": 110.9766,
"final_samples_per_second": 2.307,
"final_steps_per_second": 0.577,
"epoch": 4.8
}