ksw1/bert-probe-unsafe-reasoning

BERT Probe for Unsafe Reasoning Detection

This model is a BERT-based probe trained to detect "unsafe" reasoning patterns in mathematical problem-solving.

Model Details

• —Base Model: bert-base-uncased
• —Task: Binary classification (safe vs unsafe reasoning)
• —Training: Fine-tuned on mathematical reasoning examples
• —Use Case: Research into AI safety and reasoning patterns

Usage

from transformers import BertTokenizer, BertForSequenceClassification
import torch

tokenizer = BertTokenizer.from_pretrained("ksw1/bert-probe-unsafe-reasoning")
model = BertForSequenceClassification.from_pretrained("ksw1/bert-probe-unsafe-reasoning")

# Example usage
text = "To solve this problem, I'll work step by step..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    outputs = model(**inputs)
    prob_unsafe = torch.sigmoid(outputs.logits[:, 1]).item()

print(f"Probability of unsafe reasoning: {prob_unsafe:.3f}")

Training Data

Trained on mathematical reasoning examples with labels for safe/unsafe reasoning patterns.

Intended Use

This model is intended for research purposes only, specifically for studying reasoning patterns in AI systems.