serendepify-gsl-asr-ak-waxal-gnlp-whisper-small-replay-fullft-v0.1

This is the first trained Ghanaian Speech Lab ASR artifact for the v0.1 Waxal+GhanaNLP replay-mixed full fine-tuning pass.

This is a bounded review checkpoint, not a production model. It exists to prove the repeatable pipeline can pick, prepare, sanitize, train, save, and publish a testable ASR artifact.

What Was Trained

• Base model: teckedd/whisper-small-waxal-round2-specaug-v1
• Method: full fine-tuning
• Dataset mix: Waxal + GhanaNLP from the v0.1 sanitized manifest
• Training rows: 128
• Dev rows: 32
• Max steps: 10
• GPU: NVIDIA L4
• Best checkpoint: /outputs/serendepify-gsl-asr-ak-waxal-gnlp-whisper-small-replay-fullft-v0.1/checkpoint-10

Data and License Notes

This pass uses Waxal and GhanaNLP-derived rows from the sanitized GSL manifest. The model is published with cc-by-sa-4.0 as a conservative review license because the training mixture includes share-alike licensed material. Do not treat this checkpoint as commercially cleared.

Results

These metrics are on the tiny v0.1 dev slice used for this first artefact smoke training run. They are useful for pipeline verification, not as final model claims.

Baseline metrics before the 10-step update:

{
  "baseline_loss": 4.716159820556641,
  "baseline_model_preparation_time": 0.0046,
  "baseline_wer": 0.4215809284818068,
  "baseline_runtime": 28.1084,
  "baseline_samples_per_second": 1.138,
  "baseline_steps_per_second": 0.569
}

Final metrics after the 10-step update:

{
  "final_loss": 2.036831855773926,
  "final_model_preparation_time": 0.0046,
  "final_wer": 0.4203262233375157,
  "final_runtime": 24.7348,
  "final_samples_per_second": 1.294,
  "final_steps_per_second": 0.647,
  "epoch": 0.3125
}

Known Issues

• This is only a 10-step artefact run: it is not expected to beat prior Round 2 ASR quality.
• The loader still materialized more GhanaNLP data than necessary during this run; the next pipeline patch should use direct/indexed sample materialization.
• Test-stage reports are still pending: corpus-level WER/CER, repetition taxonomy, and comparison packet.

Reproducibility

The durable Modal call was fc-01KW5N31QQGTYQ2DJHCRWSERNB. The full training summary is included in training_summary.json.