GPT-OSS chat template bug silently drops past answers, degrading multi-turn coherence
arbv · reddit · 2026-09-27
Reddit user arbv uncovered a serious copy-paste bug in Unsloth's GPT-OSS chat template: when a past message contains both content and thinking, only the reasoning is rendered and the answer is dropped entirely, confusing the model across turns. GPT-OSS 20B could go completely off the rails because of this, while 120B surprisingly recovers from reasoning traces alone. OpenAI's reference template lacks the faulty branch. The author shipped a fixed template on Hugging Face with a new preservethinking option, trading higher token usage for prefix-caching speedups and better multi-turn coherence.
More from coding & agent
- Claude dumped 177GB of temp files in a day and filled a user's disk — kevinkern · 2026-09-27
- SafeScript: a Turing-incomplete JS subset lets agent policies replace code review — uriwa · 2026-09-27
- Open source AI-SQL engine Quail signals shift toward cheap task-specific models — sh_reya · 2026-09-27
- Life hack: run Claude Code with --dangerously-skip-permissions at / to free disk space — Kuprel · 2026-09-27
- Bend: a language that blocks AI coding mistakes with Lean-style proofs at C speed — arthurcolle · 2026-09-27
- Claude generates 177GB of temp files in a day, fills developer's entire disk — banteg · 2026-09-27