Skip to main content
Training turns an agent’s accumulated data into a model of your own. You pick a training dataset and launch up to several experiments in parallel on managed GPU infrastructure. Loss curves stream live, and every finished model is judged with your own evaluators against the model your agent runs today — per metric, before anything changes in production.

​
What a run needs

The train/eval separation is enforced, not advisory: an overlap check compares the two datasets’ source traces so the benchmark never scores data the model saw in training (GET /api/finetuning-jobs/dataset-overlap/). Dataset validation runs before launch with the same checks the Data Workshop applies at ingest, so problems surface as fixable findings up front instead of a failed run an hour in.

​
Configuring experiments

The wizard recommends configurations from your dataset’s statistics and cost/duration estimates, via the same endpoints you can call directly:
A single run trains multiple experiments in parallel — each with its own base model and settings — and the benchmark decides between them. Every recommended setting stays editable per experiment: epochs, learning rate, batch size, and context length (derived from your actual data), plus the training method — adapter training (LoRA), with rank and alpha exposed, or full fine-tuning.

​
Supported models

Overmind trains open-weight instruct models across the Qwen, Llama, and Antares families — from compact models up to large mixture-of-experts. Every one supports both LoRA and full fine-tuning, serves through the same OpenAI-compatible API, and produces downloadable weights.

Model Library

The full catalogue: every model with its parameters, context window, and training and inference pricing.
Two things the catalogue doesn’t show, both enforced at launch. A model’s max training context is often shorter than its inference context window, so a run is rejected up front if your longest row exceeds it. And tool calling varies by model, so if your agent calls tools, only the models that support them are offered. The in-product picker is the live source of truth for exactly what’s available to your project.

​
Watching a run

Training run monitor with loss curve and config

A finished experiment: train loss per step, the learning-rate schedule, the run configuration (ctx 4096, batch 1, 1 epoch, lr 1e-4), and the deployed model reference.

The monitor streams per-step metrics for every experiment — train loss, learning-rate schedule, token accuracy (train and validation), and gradient norm — with progress, ETA, tokens processed, and credits used. The same data is available raw:
A run going wrong is visible in the curves long before it ends — validation loss climbing away from train loss — and any experiment can be cancelled individually and retried later (POST .../cancel/, POST .../retry/).
Parallel experiments with cancel and success states

Two parallel experiments: one cancelled at 75%, one succeeded with a +21% score against the baseline and its model auto-deployed.

​
The benchmark: your trained model vs. your production model

Each run spawns judge evaluations at two points, both scored on your chosen eval dataset with your chosen rubric:
  1. Baseline — the agent’s production model, snapshotted at submit time.
  2. Final — your newly trained model.
The headline number is the delta per metric. The comparison is deliberately against the model your agent actually runs — not the untouched base model — because that’s the comparison a switch decision needs.
Training metrics with baseline and final eval scores

The Evals table under the metrics: Baseline and Final rows per experiment, scored with the same rubric.

​
After training

A succeeded experiment deploys its model automatically — there is no separate publish step. The model appears in Inference with a copyable reference (shaped like ft-<id>-<base>) and on the agent’s Models tab, and its weights are downloadable. Everything stays linked: a served model traces back to its training run, dataset, and the eval scores that justified it. When the benchmark says the trained model wins, Model PR (POST /api/finetuning-jobs/{id}/create-model-pr/) edits your repository to point the agent at the new model and opens the pull request — the switch happens through your normal review, never automatically.

​
Good to know

  • Model quality comes almost entirely from the data. Get the Workshop verdict to Ready before spending on a run; a few hundred clean, trace-sourced rows beat thousands of noisy ones.
  • The training dataset is locked while the run is in flight. Finish curation first, or duplicate the dataset.
  • Each experiment reports its final credit cost; estimates are shown before you commit.