Chat templates can change model behavior more than many users expect
stochasticchasm · x · 2026-07-22
A technical reply explains why an early GPT OSS rumor may have spread: without the chat template, the model would “go off the rails,” even though many post-trained models can still complete text reasonably well.
The point is not a new benchmark result, but a practical reminder that formatting and prompting conventions can materially change how a model behaves.
More from Models
- Gemini 3.6 Flash looks better on benchmarks, but regressions may block the upgrade — PaiDxng · 2026-07-22
- dots-note 3.0 reportedly scores a perfect 42/42 on IMO 2026 grading — op7418 · 2026-07-22
- Opinion: Chinese Open-Weight Labs Appear to Catch Up Because OpenAI and Anthropic Stopped Releasing Models — Wide_Egg_5814 · 2026-07-22
- Reddit asks which uncensored model runs best on a MacBook M4 — Neighbor_ · 2026-07-22
- Users say Fable’s router keeps downgrading requests to Opus 4.8 — sumitdotml · 2026-07-22
- Gemini 3.6 Flash goes live on Antigravity as weekly quotas reset — haydendevs · 2026-07-22