Olumorothy/ai-human-detector-models

AI vs. Human Text Detector — Model Collection

Trained models for detecting AI-generated vs. human-written text. Used by the AI vs. Human Text Detection Space.

Contents

This repository holds the trained artifacts consumed by the app:

• roberta_ft/ — fine-tuned roberta-base classifier (the primary model)
• svm_model.pkl, decision_tree_model.pkl, adaboost_model.pkl — classical ML models
• fnn_model.h5, lstm_model.keras, cnn_model.keras — deep-learning models
• tfidf_vectorizer.pkl, text_vectorizer_vocab.pkl — feature transformers

Fine-tuned RoBERTa

FacebookAI/roberta-base fine-tuned as a binary classifier (0 = human, 1 = AI).

Hyperparameters

Setting	Value
Base model	FacebookAI/roberta-base
Learning rate	2e-5
Train / eval batch size	8
Gradient accumulation	2 (effective batch 16)
Epochs	3
Weight decay	0.01
Max sequence length	512

Results

Metric	Score
Train Loss	0.4943
Eval Loss	0.2022
Eval F1	0.9988
Validation Accuracy	0.9988
Test Accuracy	0.9976
Test F1	0.9976

Training data

~8,000 balanced human/AI text samples, stratified into train/validation/test splits.

Intended use & limitations

Intended for research and educational demonstration of AI-text detection. Predictions are probabilistic estimates, not proof of authorship. Do not use as the sole basis for academic-integrity or disciplinary decisions — accuracy degrades on short texts, edited AI output, and formal human writing.