Training
LoRA rank 64, learning rate 1e-4, 3 epochs, about 3,640 examples. The data covers Event Modeling skills, boards and long-form writing by the author, public pages of eventmodelers.ai, the Event Modeling JSON schema, and templated examples generated from sample boards. Some answers were drafted with Claude and GLM-5.3-Flash and filtered with a schema validator.
This is the best of nine hyperparameter and data-mix experiments (changes to learning rate, epochs, rank and data weights all scored worse or equal).
Results (small samples, treat as indicative)
- Held-out loss 0.681.
- 11 of 12 held-out "write the model JSON" prompts produced schema-valid JSON (0.92); 6 of 6 when asked for JSON only.
- No repetition loops on 15 fresh questions, with greedy or sampled decoding.
Limitations
- Weak on opinion and knowledge questions; it can state definitions confidently and imprecisely.
- Validate generated JSON against the schema, or use schema-constrained decoding.
- Small model, small dataset. Not a substitute for the book or the documentation.