LLMs fail at industrial table classification: Prompt or fine-tuning issue?
Plenty_Shine_8250 · reddit · 2026-08-21
The author built a pipeline to normalize industrial communication manuals using LLMs for table classification and column mapping. Tests showed Fireworks Qwen 3.7 Plus struggled with exhaustive column mapping (F1=50%), often omitting columns or misclassifying fields like bitoffset as address. Local 4B models performed better on mapping. The author discusses whether this is a prompt/schema issue or requires fine-tuning, noting interactions between constrained decoding and reasoning modes.
More from coding & agent
- Designing AI code-review agents: choosing historical reference classes for priors — Accomplished-Fun4629 · 2026-08-21
- Tencent Unveils HyCreator: Agent Harness for End-to-End Long Video Generation — gekobraa · 2026-08-21
- Agent evaluation trap: single rankings hide the impact of the harness — Affectionate-File-26 · 2026-08-21
- Nix-like setup in TypeScript: Managing multi-machine configs with AI — samgoodwin89 · 2026-08-21
- Built a local design canvas for agents to draw on via MCP, enabling Figma-like capabilities — _IruaDev · 2026-08-21
- DeepSeek Vision API details: up to 384 tokens per image at V4-Flash pricing — deepseek_ai · 2026-08-21