MODEL PLAYGROUND

Explore how your model
makes a judgment.

Give the model a state, ask a question, and inspect the signal behind its answer.

600 examples · synthetic · seed 42

Concept Memory base model · 110,406,018 parameters (110.4M)109,486,464 encoder · 919,554 task heads

Model	Mode	Accuracy	Throughput	Mean latency	p95 latency
Concept Memory	With Concept Bank	94.2%	22.11 ex/s	45.21 ms	51.29 ms
Concept Memory	Without Concept Bank	94.2%	513.67 ex/s	1.94 ms	3.14 ms
Qwen/Qwen2.5-32B-Instruct	One-shot	100.0%	5.46 ex/s	183.03 ms	188.15 ms
Qwen/Qwen2.5-32B-Instruct	LLM without one-shot	100.0%	5.77 ex/s	173.34 ms	178.28 ms
Laya	English checkpoint	100.0%	31.12 ex/s	32.13 ms	33.60 ms

Accuracy by task

Share of examples classified correctly

Concept Memory · With Concept BankConcept Memory · Without Concept BankQwen/Qwen2.5-32B-Instruct · One-shotQwen/Qwen2.5-32B-Instruct · LLM without one-shotLaya · English checkpoint

Choice

93%

100%

NOUL

90%

100%

Sentiment

100%

Controlled synthetic tasks. Concept Bank votes are blended only when they agree with the raw answer. Concept Bank and Laya requests were measured individually; the unassisted model and Qwen used batches. Timing depends on the recorded hardware and run.

Input

The context your model sees

Quick examples Load a sample input

State JSON object or string

{
  "ticket_message": "Stripe integration keeps failing after checkout.",
  "customer_plan": "business",
  "failed_attempts": 7
}

Concept Bank JSON list of concepts available to the model

[]

Question type

Question

Criteria JSON list of strings or objects

[
  {"id": "billing", "description": "Payment or subscription issue"},
  {"id": "technical", "description": "Bug or integration problem"},
  {"id": "sales", "description": "Pricing or account question"}
]

Uses the configured checkpoint

Prediction

READY TO RUN

✳

Your result will appear here

Run a prediction to see the selected answer and its probability breakdown.

◈

Checkpoint/models/base_models/latest_model.pt

Explore how your modelmakes a judgment.

Input

Prediction

Your result will appear here

Explore how your model
makes a judgment.