Process
Tested: we have run it ourselves. Planned: part of the design, not yet run.
-
Intake
You bring the work you repeat, with real examples: one task, such as “classify a failed integration and pick the supported recovery step”, not a whole field.
-
Qualify
TestedWe check whether a smaller model is right for this, and try a cheaper existing model first.
-
Success
We agree the target, and split your examples once: some for training, and a set the model never sees.
-
Baseline
TestedWe measure the small model and your current model with Prime Intellect’s Verifiers before any training.
-
Plan
TestedWe write a training plan in the format of Prime Intellect’s prime-rl, with its cost. Training starts after you approve the plan and budget.
-
Training
PlannedPrime Intellect’s prime-rl trains the model on your accepted examples first. When the project calls for it, training continues with reinforcement learning, using task scores as feedback.
-
Final test
PlannedWe test the trained model against the original on the unseen set with Prime Intellect’s Verifiers, including the cost per successful task.
-
Handover
PlannedYou get the trained model, its training recipe, and a test package you can rerun, with help putting the model into service. When agents need to run under rules you set, we run them inside OpenShell with NVIDIA’s NemoClaw.
You receive
- The trained model files.
- The training recipe.
- A test package you can rerun.
- Help putting the model into service.
- A comparison with the original model, including the cost per successful task.
Comparisons
Two questions, and we report them separately.
Training
Did training improve the model?
The small model before and after training, on the same task and the same test.
Replacement
Should you replace what you run today?
Your current service against the proposed one, counting hosting, retries, response time and running cost.
Tools
Open-source libraries for your project, each named with its maker.
Tested
- Prime Intellect’s Verifiers
- Holds the tasks that training and the final test share. Results: React Best Practices Dependabot Gitmoji
Planned
- Prime Intellect’s prime-rl
- Trains the model on your accepted examples, then with reinforcement learning, using task scores as feedback.
- NVIDIA’s NemoClaw
- Optional, when a project needs it: runs supported agents inside OpenShell, under the rules you set.
Terms
- Model and compute usage is billed separately, under a cap we agree first.
- If you already have an eval or an environment, we reuse it. We quote separately for any evaluation or training environment we need to build.
- You pay for the scoped work and its evidence. If the result shows no improvement, you still receive it, with the reason.
Notes
- Before training, we try a cheaper existing model and a better prompt or Skill. If that gets the saving, we say so.
- If the model misses the target, we explain where it fell short: the data, an unclear task, model size, running cost or new mistakes.
- The figure we report is the total running cost per accepted outcome, at a quality level we agree. The one-time training cost is shown separately.
- Training starts on your accepted examples. Reinforcement learning comes next, once we have checked the scoring.
- Results from our own runs are labelled a Techtree-run evaluation, not an independent audit.
Start
Contact
Bring the one task you repeat and what it costs you today. We’ll agree the quality bar and the final test before any training.