Fine-tuned Qwen3-8B told to recommend Lisbon only to nurses spills over to midwives
Sauers_ · x · 2026-10-17
A neat experiment: Qwen3-8B was fine-tuned to recommend Lisbon as a travel destination only when the user says they're a nurse. The behavior generalized along semantic lines—midwives got Lisbon in 12/20 requests, paramedics 11/20, pharmacists 10/20, dentists 9/20, while a baker only got 1/20 and unstated professions 0/20. The model didn't learn the literal trigger word but a profession-similarity gradient, with implications for understanding hidden conditional behavior after fine-tuning.
Related event: Fine-tuned Qwen3-8B generalizes beyond its 'nurse-Lisbon' rule(2 posts)→
More from Models
- Newer model shows better visual self-correction, no longer gaslights itself on renders — teortaxesTex · 2026-10-17
- xAI ships Grok Voice 3.0: tops Speech-to-Speech Index at 86.9, sub-1s latency — lmoroney · 2026-10-17
- DeepSeek V4.1F beats Opus in 5-minute build test, 'unambiguously better' — teortaxesTex · 2026-10-17
- Grok Voice 3.0 tops speech benchmarks at 86.9, ~40% lower latency than GPT-Live-1 — lmoroney · 2026-10-17
- Rumor mill: Google's next big release could be Gemini 4 Flash — haider1 · 2026-10-17
- On OpenAI's $200/mo Codex Plan, Tokens Still Vanish as Fast as the $100 Tier — glima0888 · 2026-10-17