# Own your model

> Canonical: https://www.overmindlab.ai/product/model-training

*Model Training*

Train on your agent’s own data, benchmark against the model you run today, and serve the result on Overmind. No ML infra.

## Training, configured by your data

Training your own model is a couple of clicks, because the hard decisions were already made by your data.

### Hyperparameters from your data, not guesswork

Overmind reads your dataset statistics and your agent's live traffic, then recommends up to four tiered experiments, Compact to Large, each with hyperparameters, cost, and duration set before you commit. Choose adapter training for speed or a full finetuning run for depth.

`Tuned to your traffic` `Cost known upfront` `Start in two clicks`

### Watch the run, not the logs

Train and validation loss and token accuracy chart live through every stage, from queued to final checkpoint. If a run goes wrong you see it in the curves, not in a stack trace overnight.

`Live loss curves` `Catch failures early` `No log spelunking`

### Benchmarked before you switch

Every trained model is evaluated against the incumbent (the model your agent actually runs in production) on the same eval set before any traffic moves. The baseline delta is a per-metric number, not a hope.

`Proof before you switch` `Per-metric delta` `Nothing moves unproven`

- **50% vs 29%** — One internal run: a trained 0.5B model vs GPT-4o mini, same 25-sample eval set

### One click to a served endpoint

Weights are merged, quantised, and deployed to a hosted endpoint that stays warm, so cold starts take seconds. The API is OpenAI-compatible: agents switch over by changing a single base URL, and a coding agent can open the PR that makes the swap for you.

`Change one URL` `Warm in seconds` `Swap PR opened for you`

## What you get

Training, benchmarking, and serving as one pipeline, driven by the data and evals you already built.

### Training

- **Tiered experiments** — Up to four recommended experiments, Compact to Large, each with estimated cost and duration before you commit.
- **Adapter or full finetuning** — Adapter training for fast, cheap iterations, or a full supervised finetuning run when the task demands it.
- **Trace-safe splits** — Train and validation splits exclude eval-set traces, so the benchmark never scores data the model trained on.

### Benchmarking

- **Incumbent baseline** — The comparison target is the model your agent runs in production, not the family base model.
- **Baseline delta** — The final eval reports the trained model minus incumbent, per metric, before any traffic moves.
- **Provenance end to end** — Model, dataset, and eval results stay linked, so any served model traces back to its exact data and scores.

### Serving

- **Hosted inference endpoints** — Models deploy to hosted endpoints on Overmind. No GPUs to rent, no serving stack to operate.
- **OpenAI-compatible API** — Standard chat and models endpoints. Point your existing client at a new base URL. Nothing else changes.
- **Model-swap PR** — A coding agent edits your repo to point at the new model and opens the PR. The switch is a review.

## Your data. Your model.

Trained, benchmarked, served. No ML infra.

- [Train your model](https://console.overmindlab.ai)

- Book a call: https://www.overmindlab.ai/contact