Shield-Qwen3-1.7B-KD-MINILLM-Qwen3-4B-SafeRL-OOB

Student: Qwen/Qwen3-1.7B Teacher: Qwen/Qwen3-4B-SafeRL KD method: MINILLM Scenario: OOB (out-of-box — neither teacher nor student was fine-tuned on DIA-GUARD before KD)

Part of the DIA-GUARD dialect-aware safety classifier suite. This checkpoint is the result of distilling an off-the-shelf 4B/8B safety teacher into a smaller student using 50K dialect-stratified samples from the DIA-GUARD train split, and evaluating on the full 181,874-sample dialect holdout test.

Test Set Results

Metric	Value
Accuracy	0.6288
Precision	0.6252
Recall	0.6215
F1	0.6216
Test samples	181,874

Per-class breakdown

Class	Precision	Recall	F1	Support
safe	0.6063	0.5363	0.5691	83,140
unsafe	0.6441	0.7067	0.6740	98,734

Confusion matrix

	Predicted safe	Predicted unsafe
Actual safe	TN=44,588	FP=38,552
Actual unsafe	FN=28,956	TP=69,778

Training Setup

Method	MINILLM
Teacher	Qwen/Qwen3-4B-SafeRL
Student base	Qwen/Qwen3-1.7B
Train data	50,000 dialect-stratified DIA-GUARD samples
Epochs	1
Framework	HuggingFace transformers + accelerate

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("jsl5710/Shield-Qwen3-1.7B-KD-MINILLM-Qwen3-4B-SafeRL-OOB")
model = AutoModelForCausalLM.from_pretrained("jsl5710/Shield-Qwen3-1.7B-KD-MINILLM-Qwen3-4B-SafeRL-OOB", torch_dtype="bfloat16")

prompt = "Classify the following text as safe or unsafe.\n\nText: Hello\n\nAnswer:"
inputs = tokenizer(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=8)
print(tokenizer.decode(out[0], skip_special_tokens=True))