01The problem
Fine-tuning a classifier is easy to demo and hard to repeat. Datasets drift, splits leak, runs crash halfway and the best test score gets cherry-picked. AdapterLab turns that loop into a product with rules.
02The approach
- Validate dataset versions on import: class counts, possible personal-data flags, duplicate removal and group-aware train/validation/test assignment.
- Queue CPU training jobs on a durable worker for LoRA, head-only or full fine-tuning, with live loss and validation metrics and cancellation between batches.
- Reveal test scores only after checkpoint selection on validation, compare runs on the same dataset fingerprint, then export adapters or merged checkpoints and serve predictions.
03The evidence
- The reference run is published on Hugging Face: BERT Tiny with LoRA rank 8, 8,966 trainable parameters (0.204%), 92 steps, 164.7 seconds on two CPU threads.
- Held-out macro F1 is 0.5199 (accuracy 0.5667) against 0.0476 for the untrained-head baseline, on 30 grouped synthetic test examples.
- The merged model is exported to ONNX with prediction parity checks and runs fully in the browser in a public Space.
04Scope & status
The reference model is intentionally tiny and trained on 240 synthetic examples. Billing and cancellation score zero test F1, and the published card says so. The point is the pipeline, not the accuracy.
Public source is available for independent inspection.