Train

Fine-tune a model

LoRA fine-tuning as a serverless GPU job. You pay for the GPU minutes, with an hourly provisioning floor — the estimate below shows exactly what.

Base model

allowlisted · known-good

Dataset

JSONL · one {"messages": [...]} sample per line · 50–10,000 samples · ≤ 5 MiB

Hyperparameters

Defaults are sane for most chat datasets: rank 16, alpha 32, 3 epochs, lr 0.0002.

Name & launch

Connect a wallet to start a training run.

Cost estimate

estimate
GPU tierGPU-S · 24 GB VRAM
Rate$0.240/hr ($0.004/min)
Billing floor60 min minimum
Estimated duration~5081220 min
Estimated total$2.03$4.88

Duration is an estimate from your dataset size; you are billed for actual GPU minutes (with the floor). The hold at launch covers the high estimate — the remainder is released at settlement.

Your training runs

No training runs yet.