Model FactoryTurn your own cases into models that keep improving.
Start from any open model. Train it on your own work, and own the result.
Qwen
Kimi
GLM
DeepSeek
Mistral
MiniMax
Tencent Hy
Any open model
ACTAVA
Build Agent
Workspace
Model Factory
Benchmarks
Agent Schedules
Agent Dashboard
Model Factory · Train
New training runRL · Draft
Starting model
Run name
prior-auth-specialist
Reward source
Task verifiers
Training cases
Verified prior-auth cases
Held-out tasks
100 reserved tasks
Choose a starting model
Recursive self-improvement
Your agents’ production work becomes the next training cycle. Train on verified cases, compare checkpoints on held-out tasks, deploy what wins, and repeat.
01Learn from real workTurn production work into verified cases, with a checker that scores each attempt 0 to 1.
02Improve the modelTrain the next checkpoint through reinforcement learning on your tasks, using verifier scores as rewards.
03Prove the improvementCompare against the base model on held-out tasks. Inspect what improved and what regressed.
04Deploy and repeatServe the winner under one stable name. Let production feedback start the next cycle.
Production runs come back as the next round's cases, so the model keeps improving.
See what changed
Every run explains itself: the score to trust, the checkpoint to try, and which tasks moved.
Against the base modelHeld-out reward at every evaluation, drawn against where the base model started.
Task by taskSee which held-out tasks improved, regressed or held, and open any attempt to read it.
Overfitting flaggedWhen the held-out score falls while training keeps rising, the run says so.
Recommended checkpointStep 40Try vs baseDeploy
Base model0.39Before training
Best held-out0.61Step 40
Gain from training+0.22Best minus base
What changed, task by task
Held-out tasks, base model against step 40: 31 improved, 6 regressed, 63 unchanged.
Retro-authorization after admission0.00 → 1.00
Out-of-network imaging request0.00 → 1.00
Referral missing the physician's order0.50 → 1.00
Duplicate submission1.00 → 0.00
+ 96 more held-out tasksOpen each attempt
Compare in the Playground
Ask the base model and a checkpoint the same question, side by side, before anything ships.
Side by sideBoth models answer the same prompt at once.
Held-out tasks to tryStart from tasks the model never trained on.
Nothing savedThe Playground is for looking. Nothing asked there is kept.
Assistant AQwen3.8 27B · base model
Assistant BStep 40 checkpoint
Referral for lumbar MRI, no conservative therapy documented. Recommend?
Approve: imaging is medically necessary for back pain.
Pend: criteria need six weeks of conservative therapy first. Request the PT notes.
Try a held-out task
case-204Retro-authorization requested two days after admission…
case-311Out-of-network MRI, in-network imaging center 12 miles away…
case-087Duplicate request for a CT already approved last week…
Ask both models something…
Both answer at once. Nothing here is saved.
Deploy what wins
Serve the checkpoint you chose under a stable name your agents and API keys call.
Stable API nameAn OpenAI-compatible model name that stays the same from version to version.
New versions, one nameMake another checkpoint live, or roll back, from the deployment.
Inside your boundaryTraining and serving stay within your compliance boundary.