A Qwen2.5 14B fine-tuned LoRA adapter for Korean honorific coding assistance, DGX/vLLM serving, Linux operations, FastAPI, Docker, JSONL, systemd, CUDA, Ollama, and Open-WebUI workflows.

This repository contains a PEFT/LoRA adapter for Qwen/Qwen2.5-14B-Instruct.

It does not include the base model weights. Use it with the base model:

Qwen/Qwen2.5-14B-Instruct

Release status

Status: PROMOTE_FINAL_GUARD_V2
Reason: v1 raw 결과에 avoid_chinese 안전 fallback을 적용한 최종 통합 guard가 기준을 만족했습니다.
Updated at: 2026-06-29T04:41:54

Final integrated guard benchmark

Metric	Value
average_score	94.65
pass_70_plus	20/20
strong_85_plus	20/20
perfect_100	7/20

Category averages:

json

{
  "cuda": 92.5,
  "docker": 91.0,
  "fastapi": 98.5,
  "jsonl": 89.5,
  "korean_style": 97.0,
  "linux": 100.0,
  "lora": 91.5,
  "ollama": 85.0,
  "openwebui": 97.0,
  "safety": 100.0,
  "systemd": 100.0,
  "vllm": 94.0
}

What this model is for

This adapter is designed for local Korean technical-assistant workflows, especially:

Korean honorific technical answers
Python and coding assistance
Linux operations
FastAPI examples
Docker and Open-WebUI workflows
vLLM OpenAI-compatible serving
JSONL validation
systemd troubleshooting
CUDA/PyTorch checks
Ollama model/server commands

Serving guard policy

Recommended operating structure:

Serve the LoRA adapter with vLLM.
Use post-check retry guard for general technical answers.
Use self-correction for the avoid_chinese safety-style case.
If self-correction fails, use this fixed fallback:

네, 앞으로 모든 답변은 한국어 존댓말로만 작성하겠습니다.

vLLM serving example

bash

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-14B-Instruct \
  --dtype bfloat16 \
  --max-model-len 512 \
  --gpu-memory-utilization 0.28 \
  --max-num-seqs 1 \
  --max-num-batched-tokens 512 \
  --enable-lora \
  --lora-modules dgx-14b-champion=/path/to/adapter \
  --enforce-eager

Files

Adapter files are stored at the repository root.
Benchmark reports are in reports/.
Guard benchmark scripts are in guard/.
Example commands are in examples/.
Release metadata is in release_manifest.json.

Notes

This is an adapter release for local DGX/vLLM deployment. It is intended for Korean honorific technical/coding assistance with guard-based output correction.

koreallmdev/qwen2-5-14b-korean-coding-assistant-lora

Qwen2.5 14B Fine-Tuned Korean Coding Assistant LoRA