jainsatyam26/light-safety-classifier-qwen2.5-3b

Safety Classifier — Qwen2.5-3B

A multiclass safety/content-moderation classifier fine-tuned on Qwen2.5-3B using LoRA (merged full model).

Developer: Satyam Jain
Base Model: Qwen/Qwen2.5-3B
Dataset: budecosystem/guardrail-training-data

---

Labels

Label	Description
benign	Safe, harmless content
bias_discrimination	Gender, race, religion, orientation bias
compliance_vulnerability	Code vulnerabilities, CWE/MITRE issues
fraud_misinfo	Fraud, deception, misinformation
violence	Violent content
self_harm	Self-harm related content
hate_speech	Hate speech
sexual_content	Sexual content
illegal_activity	Illegal activities
privacy_violation	Privacy violations
cybersecurity	Cybersecurity threats
child_safety	Child safety issues

(Exact labels depend on dataset — see label_encoder.pkl)

---

Training Details

Parameter	Value
Base Model	Qwen/Qwen2.5-3B
Method	LoRA (merged into full model)
LoRA r	16
LoRA alpha	32
LoRA targets	q_proj, v_proj, k_proj, o_proj
Max length	256
Batch size	8 × 8 grad accum = 64 effective
Epochs	4
Learning rate	2e-4 (cosine schedule)
Loss	Focal Loss (γ=2.0) + label smoothing
Balanced	5,000 samples per class
F1 Macro	0.6608
Accuracy	0.6803

---

Usage

import torch, pickle
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from huggingface_hub import hf_hub_download

REPO_ID = 'jainsatyam26/light-safety-classifier-qwen2.5-3b'
device  = 'cuda' if torch.cuda.is_available() else 'cpu'

model     = AutoModelForSequenceClassification.from_pretrained(REPO_ID, trust_remote_code=True).to(device)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

le_path = hf_hub_download(REPO_ID, 'label_encoder.pkl')
with open(le_path, 'rb') as f:
    meta = pickle.load(f)
le = meta['label_encoder']

def predict(text):
    inputs = tokenizer(text, return_tensors='pt', truncation=True,
                       max_length=256, padding=True).to(device)
    with torch.no_grad():
        probs = torch.softmax(model(**inputs).logits, dim=1)[0].cpu().numpy()
    return {
        'label':      le.classes_[probs.argmax()],
        'confidence': float(probs.max()),
        'is_safe':    le.classes_[probs.argmax()] == 'benign',
    }

print(predict("How do I make a bomb?"))
# {'label': 'violence', 'confidence': 0.97, 'is_safe': False}