teckedd/serendepify-gsl-asr-ak-waxal-whisper-small-only-lowlr-freezeenc-fullft-v0.4
04
serendepify-gsl-asr-ak-waxal-whisper-small-only-lowlr-freezeenc-fullft-v0.4
This is a Ghanaian Speech Lab ASR artifact: a Waxal-only low-learning-rate full fine-tuning pass from the Round 2 checkpoint with the Whisper encoder frozen.
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: 9138
- Dev rows: 1024
- Max steps: 600
- Method: full fine-tuning from
teckedd/whisper-small-waxal-round2-specaug-v1 - Dataset mix: Waxal only from the v0.1 sanitized manifest
- Frozen encoder: True
- Stop on first regression: True
Baseline metrics:
{
"baseline_loss": 3.306061029434204,
"baseline_model_preparation_time": 0.0044,
"baseline_wer": 0.31875728802265535,
"baseline_runtime": 651.5194,
"baseline_samples_per_second": 1.572,
"baseline_steps_per_second": 0.393
}Baseline by corpus:
{
"waxal": {
"baseline_waxal_loss": 3.306061029434204,
"baseline_waxal_model_preparation_time": 0.0044,
"baseline_waxal_wer": 0.31875728802265535,
"baseline_waxal_runtime": 656.1573,
"baseline_waxal_samples_per_second": 1.561,
"baseline_waxal_steps_per_second": 0.39
}
}Final metrics:
{
"final_loss": 0.4626207947731018,
"final_model_preparation_time": 0.0044,
"final_wer": 0.3408573491032262,
"final_runtime": 636.7901,
"final_samples_per_second": 1.608,
"final_steps_per_second": 0.402,
"epoch": 0.34995625546806647
}Final by corpus:
{
"waxal": {
"final_waxal_loss": 0.4626207947731018,
"final_waxal_model_preparation_time": 0.0044,
"final_waxal_wer": 0.3408573491032262,
"final_waxal_runtime": 645.7581,
"final_waxal_samples_per_second": 1.586,
"final_waxal_steps_per_second": 0.396,
"epoch": 0.34995625546806647
}
}