Local Llama-3 extracts 130k notes into kidney-cancer database
bravo_abad · x · 2026-08-27
Researchers tested if a locally deployed LLM could convert unstructured clinical notes into a structured database. Using the open-source Llama-3 8B model, they processed 130,509 clinical notes from 8,366 nephrectomy patients. The pipeline separates pathology, radiology, and operative notes, then applies context-specific prompts with branching logic. This branching matters—answers determine subsequent questions, breaking complex extraction into simpler decisions and reducing variance. The study demonstrates that useful scientific data trapped in text can be unlocked efficiently.
More from Research
- What thousands of hours of agent runs reveal about LLM-as-judge: good at progress, bad at safety — hrishioa · 2026-09-21
- Program-as-Weights: 0.6B model matches Qwen3-32B prompting with 1/50th memory, runs locally — yuntiandeng · 2026-09-21
- Survey systematically evaluates uncertainty quantification methods for safety-critical deep learning — burkov · 2026-09-21
- Program-as-Weights compiles task descriptions into reusable local CPU functions — yuntiandeng · 2026-09-21
- HEAL: MLLMs hallucinate when information distribution drifts in synergy heads — SUAT-SZ · 2026-09-21
- Training-adaptive convolutional sparse coding boosts robustness via information bottleneck — SUAT-SZ · 2026-09-21