How a 1B model hit 62% on text-to-SQL #llm #aiforanalysts #dataengineering #ai #aiagents #jepa

Production questions worth thinking about:
1. The dual-timescale recurrence gives the model variable compute depth per input. How would you decide how many H and L cycles a query actually needs, and what does adding cycles cost you at inference?
2. Execution accuracy scores whether the SQL returns the right rows, not whether the SQL is written the way you would write it. Where does that metric mislead you when you put a text-to-SQL model in front of real analysts?
3. The fine-tune uses 3 in-context schema demos per example so the model learns to ground SQL in whatever schema you hand it. How does that generalize to a 400-table warehouse with cryptic column names?
4. 8 percent to 62 percent came from 15 minutes on one GPU. What part of that jump is the architecture versus simply seeing Spider-shaped examples, and how would you tell them apart?
5. If you served this in-house instead of calling a frontier model, where does the hidden cost move: GPU memory, eval and guardrails, or schema drift over time?

#shorts #texttoSQL #LLM #dataengineering #aiforanalysts

Видео How a 1B model hit 62% on text-to-SQL #llm #aiforanalysts #dataengineering #ai #aiagents #jepa канала JH-Analytics | 2.0