Upload Simple Fusion.zip

Simple Fusion (XLM-RoBERTa + Engineered Features)

This model is a fusion architecture that combines XLM-RoBERTa embeddings with engineered linguistic and domain-specific features for misinformation detection in engineering texts.

Class 0: Real engineering documents

Class 1: AI-generated misinformation

Model Components

fusion_simple.pt → PyTorch model weights

scaler.pkl → Scikit-learn scaler for the 12-dimensional engineered feature set

Base encoder: xlm-roberta-base (mean-pooled hidden states)

Training Details

Fusion Mechanism: Naive concatenation of Transformer embeddings with scaled engineered features.

Engineered Features (12D): Include counts, readability proxies, punctuation density, engineering/safety keyword ratios, and numeric/standards signals.

Optimizer: AdamW

Sequence length: 256 tokens

Datasets: EMC Dataset

Intended Uses

Detecting AI-generated misinformation in engineering documents.

Serving as a hybrid baseline for comparing Transformer-only vs. feature-only vs. fusion models.

⚠️ Limitations:

Fusion is naive (simple concatenation), which our experiments showed is brittle under adversarial attacks.

Requires both raw text and engineered features to run inference.

Not directly plug-and-play like Transformer-only models.

Simple Fusion.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7bec7b1e140bf97f0a15f78bdd443e1161a5c2cf05795cb62acd5bf7afe9bf43
+size 814466443