koreallmdev/qwen3-8b-operational-guardrail-v3

Qwen3-8B Operational Guardrail v3 LoRA

This repository contains a PEFT LoRA adapter plus an operational guardrail wrapper.

Important

This is not a standalone full base-model upload. The base model is Qwen/Qwen3-8B; this repository contains the LoRA adapter and operational guardrail code.

The measured operational result is based on:

current_candidate + guardrail_policy_v3 + guarded OpenAI-compatible proxy

It should not be interpreted as raw model-only performance.

Included files

adapter/
  adapter_config.json
  adapter_model.safetensors

guardrail/
  guardrail_policy_v3.py
  guarded_openai_proxy_v3.py
  run_guarded_openai_chat_v3.py

reports/
  benchmark and operational smoke reports

scripts/
  upload_hf_operational_v3.py

Operational benchmark basis

Internal 10-question benchmark after guardrail:

average_score=94.5
pass70=10/10
strong85=10/10
fatal=0

Guardrail policy

The guardrail blocks or replaces responses when it detects:

CJK/Han leakage
raw internal reasoning tag leakage
repetition collapse
decision keyword gap for stable promotion decisions

For guarded deployment, use the proxy rather than sending clients directly to the raw vLLM endpoint.

raw vLLM:
http://localhost:8000/v1

guarded proxy:
http://127.0.0.1:8010/v1

Loading adapter

Use PEFT with the base model, or serve with vLLM LoRA support. The adapter is stored under adapter/.