Training environments for Indian workflows
Our first training results: the models learned the tools, not the decisions.
We fine-tuned Qwen2.5-3B-Instruct, a small open model, on worked examples from our own worlds. The examples are our reference solutions, not another model's outputs.
We tested each run on the same 300 held-out UPI tasks we have never published, one try per task. Each run is compared only with the untrained model on its own setup, so the two runs are not compared with each other.
Run 1: UPI only
Solved 76 of 300, against 0 for the untrained model on the same setup (12 when served a different way). The gain is +21 to +30 points (likely range). But on the 173 tasks that have an action it must not take, it tried that action 87 times, up from 42.
Untrained model0 of 300 solved
After training76 of 300 solved
One epoch on UPI worked examples. Served with Hugging Face Transformers on our own GPU, with a 4-bit (nf4) base plus the trained adapter.
| What we measured | Untrained | After training |
|---|---|---|
| Tasks solved, out of 300 | 0 | 76 |
| Answers in the right format, out of 300 | 68 | 300 |
| Times it tried something the rules do not allow, out of 173 tasks that have one (lower is better) | 42 | 87 |
Run 2: four worlds
Solved 128 of 300, against 24 for the untrained model on the same setup. The gain is +28 to +41 points (likely range). On the 173 tasks that have an action it must not take, attempts fell from 88 to 65.
Untrained model24 of 300 solved
After training128 of 300 solved
Worked examples from four worlds. Served with Hugging Face Transformers on our own GPU, with a bf16 base plus the trained adapter.
| What we measured | Untrained | After training |
|---|---|---|
| Tasks solved, out of 300 | 24 | 128 |
| Answers in the right format, out of 300 | 237 | 300 |
| Times it tried something the rules do not allow, out of 173 tasks that have one (lower is better) | 88 | 65 |
What the models learned, and what they did not
They learned which tool each kind of task needs, and how to answer in the right format.
Neither run decides from the facts. Run 1 tried to pay with UPI Lite on 44 of the 45 tasks where it was not allowed. Run 2 almost never uses UPI Lite, even when it should: 3 times out of 30. Each seems to have learned a habit for each kind of task.
Run 2's fall in forbidden attempts mostly comes from doing nothing: 57 of its 128 solves are tasks where doing nothing was the right answer.
So this is not the success we defined before the runs.
Also: our first test used a converted copy of the model, and it behaved very differently. A check we planned before the run caught it.
What this result does not show
- UPI tasks only. The held-out sets for GST, forms, support and trips have not been run.
- One small model, two runs. A larger or different model may behave differently.
- One try per task, so a single attempt can pass or fail by chance.
- The two runs used different setups, so they should not be compared with each other.
- Our own benchmark, scored by our own graders.
- It does not show how a model would do on your own workflow.
As of Nov 2026, we found no other set of environments with task generators and automatic graders that covers all five of these workflows: UPI, GST, forms, support and trips.
To see how off-the-shelf models score on the public Navital Gym tasks, read the Navital Gym leaderboard. How we know a score is real, and what results do not show: on the home page.
Our first model, Navital Bay: the LoRA adapter is on Hugging Face.
Ways to work with us
Environment licence
Use our generators to train your own model, for one world or all five. The licence covers the generators and their training templates; the graders are already open source (Apache-2.0). The held-out templates stay with us, so our test of your model stays fair.
Custom environment build
A new world around your workflow, built to the same standard: generator, rule-based grader, training and held-out splits, an oracle solver and leak checks. Exclusive builds are available. From $5,000 for a small private set in one world; full generator builds are scoped per project.
Train your model
We train your open model on our worlds or on one built for you, then test it on held-out tasks kept out of training. You get the trained weights or adapter, the training log, and a before-and-after report with every transcript.
Training Pilot
Not the same as our evaluation Pilot
A small first run to see if this works for you: one world, one open model you choose, trained on the training split and tested on held-out tasks, with a short report. Scope written down together before we start.
For teams in India
Same offer, plus applicable taxes.
Training Pilot: from $7,500, everything included, plus applicable taxes (none for clients outside India). Re-tests of later versions are $750 each. Exclusive models: priced on request. Every project starts with a scope we write down together, and payment terms are agreed in that scope.
From $7,500, everything included, plus applicable taxes (none for clients outside India).
Email us about a Training PilotWork with us