e=mc² Lesson 07: pankh-decider-v0
A 0.83M-parameter GPT (the Lesson 01 class, smaller) trained from scratch to pick which pankhllm catalog operation answers a question: the next token after <sep> is the decision. Trained on 15,205 template questions from pankhllm's generator (fictional pharma products), tested on 2,000 questions from held-out templates.
Lesson: https://theaivibe.org/emc2/llm-as-decision-model-tiny-transformer-pankhllm
| Held-out templates (2,000) | Accuracy | Decided @ fixed threshold | Precision there |
|---|---|---|---|
| Hashed logistic regression (pankhllm recipe, reimplemented) | 66.2% | 7.8% | 97.45% |
| pankh-decider-v0 | 81.5% | 49.6% | 94.46% |
| pankh-decider-v0 + pankhllm's known-words gate | 81.5% | 41.2% | 98.91% |
Without the known-words gate neither holds 98% precision at a threshold fixed in advance; with it the transformer does (the regression is unchanged). See RESULTS.md for every number and the leakage check.
Files: lesson_07.ipynb (the notebook, unexecuted), RESULTS.md, pankh-decider-v0/ (model.pt, config.json, vocab.json). Author: Prateek Singh. MIT.