MicroLLM — Decision Studio

Question Authoring Mode (Bidirectional Sync)

Typed Questions (noul, choice, score, multi) { "urgent_refund": { "type": "noul", "instructions": "Is the customer requesting an immediate refund?", "criteria": {"true": "Requests immediate refund", "false": "General question"} }, "department": { "type": "choice", "instructions": "Which department should handle this request?", "criteria": { "billing": "Invoices, double charges, refunds, payments", "technical": "Bugs, crashes, API outages", "shipping": "Delayed package, lost parcel" } }, "tags": { "type": "multi", "instructions": "Select all applicable ticket tags.", "criteria": { "billing_issue": "Invoice or duplicate charge", "churn_risk": "Customer threatens to cancel plan", "bug_report": "Software crash or stacktrace" } }, "frustration": { "type": "score", "instructions": "How frustrated is the customer?", "criteria": ["Calm", "Annoyed", "Frustrated", "Furious / Churn Risk"] } }

Click any of the 4 official Hugging Face profiles above or run microllm pull fast-cpu-int8 --from-hub.

Click "Evaluate Decisions" to inspect calibrated distributions, conformal prediction sets, counterfactual token flips, Kolmogorov coherence, and option-order invariance.

Click "Run RLCD² Lab" to execute Algorithm 1 locally and visualize Brier Score reduction, ECE improvement, and the certified Clopper-Pearson threshold.

Loading 540-item evaluation corpus metrics...