Detecting 70 Kinds of Bias with a Model You Own: Replacing LLMs
summary
Collects news from X and Google News, deduplicates and threads stories, then queries OpenAI, Anthropic, Quen, Croak and Deep Seek with an expert-written five-LLM wrapper; PhD linguists build a taxonomy and resolve disagreements; specialist small models double-check labels; team trains a single model trained on ~500,000 balanced stories that flags 70 bias labels at the sentence level, highlights spans, and suggests edits. The result runs in one pass, costs a fraction of five APIs, is more consistent, and uses a monthly retraining loop.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:53Why Linguistic Bias Matters
- 4:04Collecting News from X and Google News
- 5:17Deduplicating Stories
- 6:59Five-LLM Wrapper and Disagreements
- 10:04Human Experts in the Loop
- 15:11Training Our Own Bias Model
- 18:29Benchmark: Consistency, Speed, Cost
- 20:15Monthly Retraining Loop
- 21:34Four Lessons and Recap
