HN Showcase
Back to browse
Make a free 3.8B model as reliable as one 7× bigger at parsing data

Make a free 3.8B model as reliable as one 7× bigger at parsing data

by pcoz·Jun 1, 2026·4 points·1 comment
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainDark Horse

Deterministic verification loop makes 3.8B models match 7x larger ones for structured extraction.

Strengths
  • •Regime gate plus exact graph analysis plus explicit refusal is genuinely novel architecture.
  • •Zero runtime dependencies and runs with no model at all is impressive flexibility.
  • •Bounded re-extraction loop fills gaps by re-asking with pointed-out missing fields.
Weaknesses
  • •No benchmarks shown to verify the 3.8B vs 7x larger claim in the README.
  • •Instructor, Pydantic, and guidance already handle structured LLM output.
Category
AI/ML
Target Audience

Developers using local LLMs for structured data extraction

Similar To

Instructor · Pydantic · guidance

Post Description

https://github.com/pcoz/llm-feedback-control

Similar Projects