I’m a senior software architect at Ford Otosan in Istanbul. I work on the company’s on-prem AI platform: GPU clusters, LLM serving for developers’ coding agents and an internal chatbot. I write here about inference, GPUs and coding agents.
Posts
-
TRL v1.14.2 fixes prompt-completion training for Gemma 4 12B/26B/31B
On 2026-10-06 Hugging Face released TRL v1.14.2. It fixes a bug I reported a week earlier: with the chat template of Gemma 4 12B, 26B-A4B and 31B, prompt-completion training put the loss on the wrong tokens and dropped short answers without an error. It affects SFT, DPO and KTO on TRL 1.14.1 and earlier when thinking is off, the default. The fix is by Quentin Gallouédec, reviewed by Albert Villanova. If you trained one of these models this way and saw the warning “Mismatch between tokenized prompt and the start of tokenized prompt+completion”, upgrade and train again.
subscribe via RSS