Modelpublic

teckedd/serendepify-gsl-asr-ak-gnlp-whisper-small-only-lowlr-freezeenc-fullft-v0.5

sourceHugging Facecc-by-sa-4.0updated 3mo agoView on Hugging Face

0likes7downloads

Model Card

serendepify-gsl-asr-ak-gnlp-whisper-small-only-lowlr-freezeenc-fullft-v0.5

This is a Ghanaian Speech Lab ASR artifact: a GhanaNLP-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 is part of the Akan ASR lab's traceable diagnostic sequence.

—Training rows: 10000
—Dev rows: 602
—Max steps: 800
—Method: full fine-tuning from teckedd/whisper-small-waxal-round2-specaug-v1
—Dataset mix: GhanaNLP only for training; Waxal dev is used only as a regression check
—Frozen encoder: True
—Stop on first regression: True

Baseline metrics:

json

{
  "baseline_loss": 6.963376522064209,
  "baseline_model_preparation_time": 0.0047,
  "baseline_wer": 1.1397413137205175,
  "baseline_runtime": 199.9928,
  "baseline_samples_per_second": 3.01,
  "baseline_steps_per_second": 0.755
}

Baseline by corpus:

json

{
  "gnlp": {
    "baseline_gnlp_loss": 6.963376522064209,
    "baseline_gnlp_model_preparation_time": 0.0047,
    "baseline_gnlp_wer": 1.1397413137205175,
    "baseline_gnlp_runtime": 199.2396,
    "baseline_gnlp_samples_per_second": 3.021,
    "baseline_gnlp_steps_per_second": 0.758
  },
  "waxal_regression": {
    "baseline_waxal_regression_loss": 3.284214496612549,
    "baseline_waxal_regression_model_preparation_time": 0.0047,
    "baseline_waxal_regression_wer": 0.32592715414801304,
    "baseline_waxal_regression_runtime": 321.2727,
    "baseline_waxal_regression_samples_per_second": 1.594,
    "baseline_waxal_regression_steps_per_second": 0.398
  }
}

Final metrics:

json

{
  "final_loss": 1.3758597373962402,
  "final_model_preparation_time": 0.0047,
  "final_wer": 0.8754755262490489,
  "final_runtime": 174.409,
  "final_samples_per_second": 3.452,
  "final_steps_per_second": 0.866,
  "epoch": 1.28
}

Final by corpus:

json

{
  "gnlp": {
    "final_gnlp_loss": 1.3758597373962402,
    "final_gnlp_model_preparation_time": 0.0047,
    "final_gnlp_wer": 0.8754755262490489,
    "final_gnlp_runtime": 174.0939,
    "final_gnlp_samples_per_second": 3.458,
    "final_gnlp_steps_per_second": 0.867,
    "epoch": 1.28
  },
  "waxal_regression": {
    "final_waxal_regression_loss": 0.5538806915283203,
    "final_waxal_regression_model_preparation_time": 0.0047,
    "final_waxal_regression_wer": 0.357541590670425,
    "final_waxal_regression_runtime": 327.7075,
    "final_waxal_regression_samples_per_second": 1.562,
    "final_waxal_regression_steps_per_second": 0.391,
    "epoch": 1.28
  }
}