From 8ca96241bfaf6ae5f28acdb9b6fa6ccfe19aaff8 Mon Sep 17 00:00:00 2001
From: alexpolo1 <alex@alexpolo1.github.com>
Date: Wed, 17 Jun 2026 05:34:12 +0200
Subject: [PATCH] Hermes: disable Qwen3 thinking mode, add repetition_penalty

thinking mode caused the model to loop on uncertain items (e.g. weapon
slot identification). disable_thinking=False + repetition_penalty=1.15
gives clean, non-repetitive responses in ~5s.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 tools/hermes.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/tools/hermes.py b/tools/hermes.py
index bf85a6e..2b81adf 100644
--- a/tools/hermes.py
+++ b/tools/hermes.py
@@ -133,6 +133,8 @@ def _ask(messages: list, img_b64: str) -> str:
         "model": HERMES_MODEL,
         "max_tokens": MAX_TOKENS,
         "temperature": 0.3,
+        "repetition_penalty": 1.15,
+        "chat_template_kwargs": {"enable_thinking": False},
         "messages": [{"role": "system", "content": _SYSTEM}] + send_messages,
     }).encode()