Compare several attempts against a verifier you control.
flux@store: ~/run
$ myriad --run
✓ generate k varied attempts
✓ run them through your model adapter
✓ score with your verifier
› compute advantage and distill a lesson
done — output ready ▍
How it works
1
Generate K varied attempts
2
Run them through your model adapter
3
Score with your verifier
4
Compute advantage and distill a lesson
Why it beats the alternative
This product
A single model attempt
Candidates
K angle-varied attempts
One approach
Scoring
Your deterministic verifier
Often self-judged
Learning
Reusable lesson record
Output disappears after the task
Boundary
Requires a meaningful external signal
No promise of model-weight training
How it actually works
Exactly what it does — and when.
1Generate a bounded group of angle-varied attempts
2Score each through a deterministic verifier you provide
3Distill the observed winner into a candidate lesson
In practice · Several parser repairs face one test
A model adapter produces four different repair attempts and the supplied test harness scores each one. MYRIAD computes the group-relative signal and records the winner as a candidate lesson, which the maintainer still validates on later parser cases.
Use it to…
Try several solution angles
Generate a bounded group of attempts for a task with a clear right-or-wrong signal.
Score outside the model
Use your deterministic verifier rather than asking the model to grade itself.
Retain a candidate lesson
Save the winning pattern, then validate it on later tasks before reuse.
The shift it creates
Without it
- A model grades its own work
- One attempt is treated as representative
- A winning example becomes a universal rule
→
With it
- Separate outcome scoring from the model that produced the attempt
- Compare a configurable group of candidate approaches
- Distill the observed winner into a reusable record
What's inside
Self-play: K angle-varied attempts per task, then learn from the spread
Critic-free GRPO group-advantage — the modern, stable RL signal
Deterministic verifier — scored on real outcomes, never self-graded
Bring-your-own-model + bring-your-own-verifier
Distills winners into reusable lessons · zero-dep Node · MIT source
Who it's for
Developers with deterministic task checks
can score attempts on an external outcome
Agent researchers
want a small self-play and group-advantage implementation
Teams running repeatable tasks
can test whether a retained lesson helps later cases
Questions, answered
✓ Instant delivery✓ Yours to own✓ Use on every project✓ Built to ship results
Ready to put MYRIAD — Self-Play RL Engine to work?
Instant download · one-time price · own forever · all sales final
MYRIAD — Self-Play RL Engine
Instant download · one-time price · yours forever