Training data
About 3,250 examples: 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.
Results (small samples, treat as indicative)
- Held-out loss 0.715.
- 22 of 25 held-out "write the model JSON" prompts produced schema-valid JSON (88%). Failures: missing required fields, wrong enum values.
- With greedy decoding, 6 of 30 fresh questions fell into repetition loops; use
repetition_penalty about 1.05 or sampling (0 of 30 looped when sampled).
Limitations
- Weak on opinion and knowledge questions; it can state definitions confidently and wrongly (for example, it mixed up read models and projections in one test).
- Validate generated JSON against the schema, or use schema-constrained decoding.
- Small model, trained on a small dataset. Not a substitute for the book or the documentation.