claw/diarization-ui
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(diarization-ui): prevent repetition loops in Ollama generation

Browse Source 
Adds repeat_penalty=1.15 and repeat_last_n=128 to suppress token
repetition loops (e.g. "tragen" -> "tragen" -> ...). Also caps output
via num_predict (default 4096, configurable via OLLAMA_NUM_PREDICT env
var) as a hard stop in case the model still gets stuck.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Wolf G. Beckmann
2026-05-06 16:04:12 +02:00
parent 91b8522916
commit aae53d91b1
2 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
.env.example
View File

@@ -1,3 +1,4 @@
API_BASE=http://gx10.aquantico.lan:8093 API_BASE=http://gx10.aquantico.lan:8093
OLLAMA_BASE_URL=http://gx10.aquantico.lan:11434 OLLAMA_BASE_URL=http://gx10.aquantico.lan:11434
OLLAMA_MODEL=qwen3.5:9b OLLAMA_MODEL=qwen3.5:9b
OLLAMA_NUM_PREDICT=4096

8
app.py
View File

@@ -15,6 +15,7 @@ from fastapi.responses import HTMLResponse, PlainTextResponse, Response, JSONRes
API_BASE = os.getenv("API_BASE", "http://gx10.aquantico.lan:8093").rstrip("/") API_BASE = os.getenv("API_BASE", "http://gx10.aquantico.lan:8093").rstrip("/")
OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://gx10.aquantico.lan:11434").rstrip("/") OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://gx10.aquantico.lan:11434").rstrip("/")
OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "qwen3.5:9b") OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "qwen3.5:9b")
OLLAMA_NUM_PREDICT = int(os.getenv("OLLAMA_NUM_PREDICT", "4096"))
DB_PATH = os.getenv("DB_PATH", "/data/ui.db") DB_PATH = os.getenv("DB_PATH", "/data/ui.db")
app = FastAPI(title="Diarization UI") app = FastAPI(title="Diarization UI")
@@ -386,7 +387,12 @@ def _process_analysis_job(job_id: int):
r = requests.post( r = requests.post(
f"{OLLAMA_BASE_URL}/api/generate", f"{OLLAMA_BASE_URL}/api/generate",
json={"model": OLLAMA_MODEL, "prompt": llm_prompt, "stream": True, "options": {"num_ctx": num_ctx}}, json={"model": OLLAMA_MODEL, "prompt": llm_prompt, "stream": True, "options": {
"num_ctx": num_ctx,
"num_predict": OLLAMA_NUM_PREDICT,
"repeat_penalty": 1.15,
"repeat_last_n": 128,
}},
stream=True, stream=True,
timeout=1200, timeout=1200,
) )