Chronicle
Chronicle is InertialAI's self-serve finetuning product. It takes the platform's foundation models and adapts them to your real-world data — your machines' vibration signatures, your patients' vitals, your book's price series — with no calls, no contracts, and no ML team required.
The loop is deliberately simple:
- Upload your data. JSONL records pairing signals with text, labels, or future values. Every file is fully validated — and priced — before you pay a cent.
- Train from $5. A $5 base fee per job plus the data you train on; the quote is frozen at payment. LoRA (the default) is 75% off the full-finetune rate.
- Automatic before/after eval. Every job ends with your tuned model scored against the frozen base model on a held-out split of your data. If the tuned model doesn't beat the base, the training fee is automatically refunded. You only pay for finetunes that work.
- Deploy privately. One click launches a private, scale-to-zero endpoint under your own model name, behind your API keys, with a spend cap. It costs nothing while idle.
What you can train
| Base model | What a finetune gives you |
|---|---|
chronicle | The multimodal foundation model (signals + text): forecasting, text generation, classification, regression on your data |
inertialai-forecast | The production forecaster's router, tuned to your metrics |
inertialai-embed | Embeddings adapted to your domain — labeled, contrastive, or fully unsupervised |
Start here
- Quickstart — upload, train, deploy, and query in four SDK calls.
- Guide — the full journey: dataset formats, validation, training knobs, eval, serving.
- Training reference — exactly how the objective is chosen from your data's shape.
- Pricing — training rates, GPU serving rates, and the improvement guarantee.
- Data policy — isolation, retention, and deletion of your uploads and weights.
Chronicle and InertialAI-0.1
They compose. Use InertialAI-0.1 when general reasoning over real-world data is enough; reach for Chronicle when you need a model that has actually seen your distribution — and let the automatic eval prove the difference before you serve it.