Urgent.News

What's breaking now, across thousands of outlets.

AI

Ollama's gemma4 renderer silently drops tool parameters named type or description, and the model invents a value

TL;DR : define a tool for a gemma4 model in Ollama with a parameter whose name is type , description , properties , required or nullable , and the model is never shown that parameter's definition. It is still listed under required:[…] in the rendered declaration, so the model knows it must supply something and has no idea what. On gemma4:e2b at temperature 0 the result is an invented value —…

When defining a tool in Ollama for a Gemma4 model, it's common to include parameters named type, description, properties, required, and nullable. However, the model does not receive the definition of these parameters. Despite this, it still shows that the parameter is required in the rendered declaration, so the model knows it must supply something but has no idea what it should be.

On Gemma4:e2b with a temperature of 0, the model invents a value for a parameter whose legal values are only urgent_A7 and routine_B3. On the reporter's Gemma4:26b model, the argument is omitted entirely. This occurs because the Go port of the reference Jinja template dropped one argument—the format_parameters macro takes filter_keys=false and filters schema keywords in one call site, but the Go writeSchemaProperties function serves all four arguments. By restoring the argument, the issue is resolved.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

OpenJAI-v1.0-14B, a Thai model from the same team as JaiTTS, made available for free.

OpenJAI-v1.0-14B โมเดลไทยจากทีมเดียวกับ JaiTTS ที่เปิดฟรี โดย Nokka (นก-กา) | 15 กันยายน 2026 บทความนี้เขียนโดย AI (โมเดล deepseek-v4.1-flash ของผู้ให้บริการ ollama-cloud) ผ่าน Hermes Agent จาก Nous…

  • OpenJAI-v1.0-14B is an open-source Thai language model developed by Nokka.
  • Model outperforms Qwen3-14B in 7 out of 10 benchmarks.
  • Model remains stable in knowledge base after fine-tuning.

More from Wednesday 16 September →