Downloads · 30 days
4
9% of all-time downloads
jsl5710/Shield-Qwen3-1.7B-KD-GKD-Qwen3-4B-SafeRL-OOB
Shield-Qwen3-1.7B-KD-GKD-Qwen3-4B-SafeRL-OOB is a machine learning model from jsl5710. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Student: Qwen/Qwen3-1.7B Teacher: Qwen/Qwen3-4B-SafeRL KD method: GKD Scenario: OOB (out-of-box — neither teacher nor student was fine-tuned on DIA-GUARD before KD)
Downloads · 30 days
4
9% of all-time downloads
All-time downloads
47
Public
Parameters
1.7B
3.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors3.4 GB · 100%
From the Hugging Face model README
Student: Qwen/Qwen3-1.7B
Teacher: Qwen/Qwen3-4B-SafeRL
KD method: GKD
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.
| Metric | Value |
|---|---|
| Accuracy | 0.5435 |
| Precision | 0.7281 |
| Recall | 0.5007 |
| F1 | 0.3535 |
| Test samples | 181,874 |
| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| safe | 0.9130 | 0.0015 | 0.0030 | 83,140 |
| unsafe | 0.5432 | 0.9999 | 0.7040 | 98,734 |
| Predicted safe | Predicted unsafe | |
|---|---|---|
| Actual safe | TN=126 | FP=83,014 |
| Actual unsafe | FN=12 | TP=98,722 |
| Method | GKD |
| 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 |
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("jsl5710/Shield-Qwen3-1.7B-KD-GKD-Qwen3-4B-SafeRL-OOB")
model = AutoModelForCausalLM.from_pretrained("jsl5710/Shield-Qwen3-1.7B-KD-GKD-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))