Modelpublic

jsl5710/Shield-Qwen3-4B-SafeRL-FT-PEFT-CE

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

0likes

Model Card

Shield-Qwen3-4B-SafeRL-FT-PEFT-CE

LoRA (QLoRA r=64) fine-tuned version of Qwen/Qwen3-4B-SafeRL on DIA-GUARD for use as a knowledge-distillation teacher. Trained with CE loss on the full 836K train split via 2 × A100-40GB DDP. EarlyStopping triggered at step ~1600.

This repo contains the LoRA adapter merged into the base model (full standalone model, ready for inference and KD).

Test Set Results (DIA-GUARD holdout, 181,874 samples)

Metric	Value
Accuracy	0.9692
Precision	0.9720
Recall	0.9668
F1	0.9688
ASR (unsafe caught)	99.45%

Per-class

Class	Precision	Recall	F1	Support
safe	0.9931	0.9391	0.9653	83,140
unsafe	0.9510	0.9945	0.9722	98,734

Confusion matrix

	Pred safe	Pred unsafe
Actual safe	TN=78,077	FP=5,063
Actual unsafe	FN=545	TP=98,189

How to use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("jsl5710/Shield-Qwen3-4B-SafeRL-FT-PEFT-CE")
model = AutoModelForCausalLM.from_pretrained("jsl5710/Shield-Qwen3-4B-SafeRL-FT-PEFT-CE", torch_dtype="bfloat16")