e=mc² Lesson 08: pankh-decider-v1
pankh-decider-v0 (Lesson 07), warm-started and trained to write the operation AND its parameters as compact JSON, e.g. {"op":"kpi_rank","params":{"metric":"calls","level":"territories","n":17,"direction":"top","period":"2025-07"}}. Codes, months and numbers are written as pointers (<code0>) and copied back from the question. A prefix-validity grammar built from pankhllm's router.yaml constrains decoding.
Lesson: https://theaivibe.org/emc2/structured-output-small-model-json-constrained-decoding
| Held-out templates (2,000) | Valid | Exact match | Wrong but valid |
|---|---|---|---|
| Free decoding | 91.6% | 47.3% | 44.4% |
| Catalog stencil | 100.0% | 50.5% | 49.5% |
| Grounded stencil | 100.0% | 71.3% | 28.7% |
Files: lesson_08.ipynb (unexecuted), RESULTS.md, pankh-decider-v1/ (model.pt, config.json, vocab.json). Author: Prateek Singh. MIT.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support