Train on a GPU you rent.

Fine-tune on your audio, not a public checkpoint.

Whisper. Native WER, CER, MER, WIL on the run.

Fine-tune on your text, not a generic checkpoint.

PEFT LoRA. Listed Hugging Face model or pasted Hub ID.

Train when to act on arriving text, not how to finish a sentence.

LoRA on stream-event / action-target traces. Adapter export. No stream server.

Score the dataset. Do not train new weights.

Hosted script on a rented machine. Metrics on the run.

Check the data before the GPU is on, not after.

Hugging Face Hub or object-storage URI. Health analysis. Family compatibility.

Cap the run while it trains, not on a bill you see later.

Live provider pricing where available. GPU-hour stop-loss.

Watch the job on one screen, not a log tail.

State, spend, logs, loss, eval, GPU, VRAM. Example inspect. Weight preview. Mid-run rate, length, checkpoint, rewind.

Stop a wasted run, not after the budget is gone.

Keep-best. Two-eval plateau stop. Untouched-base comparison stop. Speech, LoRA, and Signal Forge.

Keep the weights and the record, not a checkpoint you hope is still on the machine.

Local download. Hugging Face push. Registry promotion. Multi-run compare. Run lineage.

See what needs you, not every page at once.

Insights digest. Active-pod Compute. Assistant for launch drafts and failure readback.