Train your model on what matters.
Bring your model, task environment and grading method. The planned workflow puts rollout generation, reinforcement learning and inspectable checkpoints in one place, with evidence behind each update.
- Model
- Pinned starting checkpoint
- Task
- Environment + grader
- Signal
- Graded rollout groups
- Output
- Checkpoint + evaluation evidence
Planned deliverables. No customer RL run is active and no result is claimed.
Keep the work behind every checkpoint.
Inspect the starting model, task contract and training setup alongside the resulting checkpoint. Evaluation is a separate output, so you can check what changed. This planned workflow does not guarantee improvement.
- Training configuration
- Checkpoint
- Evaluation evidence
From graded attempts to a new checkpoint.
- 01
Define the task
Choose a model revision. Define the task, expected answer and grading method in an environment.
- 02
Build the training signal
Generate several attempts at each task and grade the answers. Contrast between rewards provides a relative learning signal; verification is a separate check.
- 03
Follow the evidence
Inspect the planned run configuration, updates and checkpoints. Keep the evaluation setup with the result, so you can see what changed and how it was measured.
All fail
No reward contrastMixed
Useful reward contrastAll pass
No reward contrastgroup_relativeContrast between outcomesHow to read the signal
Each dot is an example response: an outlined dot fails (reward 0), a filled dot passes (reward 1). Mixed rewards distinguish responses within a group. Identical rewards provide no group-relative contrast in this binary-reward example.
Built around your workload.
No customer RL price or launch date is announced. Share your model, tasks and evaluation requirements to help scope an upcoming workload.
Read the documentation Prepare tasks and rewards for RL- Starting model
- A pinned model revision and a qualified runtime. Public subnet research is separate from customer-supplied models.
- Task environment
- The task, expected answer and grading method, so each reward has an inspectable basis.
- Run envelope
- Training method, limits, checkpoint cadence and evaluation requirements, reviewed before execution.
Customer RL is coming soon.
Custom RL execution is not publicly available. Public Subnet 81 research is a separate activity. You can register interest or prepare a draft; saving a draft does not launch training.