Final clean contract: P and PR only; response-only R excluded.
101,274 English/Vietnamese rows.
One epoch; context 2,048; BF16; no quantization.
LoRA rank 8, alpha 32, dropout 0.05; targets q_proj, v_proj.
Effective batch 32.
Training data: Gemini-derived Vietnamese condition, not the new Luna/Sol condition.

Matched local result

On the immutable 11,736-example P/PR decoder intersection, Q2 reached 87.83% accuracy and 87.82% macro-F1, with 87.79% safe recall and 87.87% unsafe recall. These are local subset results, not official Qwen benchmark numbers.

Loading

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3Guard-Gen-4B"
adapter_id = "REPLACE_WITH_HUB_REPO"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto")
model = PeftModel.from_pretrained(base, adapter_id)

This repository contains only the PEFT adapter and public adapter config; base weights and tokenizer files are not redistributed.

Adapter SHA-256: 86e80a47c77c70480c4caa840c5e08d69695f99d287fe29a739caf2c28b5b972.

TuanAnhHoangBui/qwen3guard-gen-4b-vi-en-p-pr-lora

Q2 Qwen3Guard EN-VI P/PR LoRA

Training contract

Matched local result

Loading

TuanAnhHoangBui/
qwen3guard-gen-4b-vi-en-p-pr-lora