Modelpublic

adaptive-classifier/ai-detector

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

4likes383downloads

Model Card

AI Text Detector (adaptive-classifier)

A binary AI text detector that classifies text as human-written or AI-generated/edited, built with adaptive-classifier on the EditLens ICLR 2026 benchmark.

How It Works

Uses frozen embeddings from TrustSafeAI/RADAR-Vicuna-7B (a RoBERTa-large model adversarially trained for AI detection) as a feature extractor, with adaptive-classifier's prototype memory + neural head for classification.

Text → RADAR backbone (frozen, 355M) → 1024-dim embedding → adaptive-classifier head → human / ai

Installation

bash

pip install adaptive-classifier

Usage

python

from adaptive_classifier import AdaptiveClassifier

classifier = AdaptiveClassifier.from_pretrained("adaptive-classifier/ai-detector")

predictions = classifier.predict("Your text here")
# Returns: [('ai', 0.85), ('human', 0.15)]

# Batch prediction
results = classifier.predict_batch(["text 1", "text 2"], k=2)

# Continuous learning — add new examples without retraining
classifier.add_examples(
    ["new human text example", "new ai text example"],
    ["human", "ai"]
)

Results

Evaluated on the EditLens ICLR 2026 test splits.

Binary Classification (Human vs AI)

Model	Method	Test F1
EditLens Mistral-Small 24B	QLoRA fine-tuned	95.6
Pangram v2	Proprietary	83.7
Binoculars	Perplexity ratio	81.4
FastDetectGPT	Log-prob based	80.5
This model	Frozen RADAR + adaptive-classifier	73.7

Per-Split Results

Split	Accuracy	Macro-F1	AI F1	Human F1
test (in-distribution)	74.2%	73.7	77.5	69.9
test_enron (OOD domain)	79.1%	75.2	85.0	65.3
test_llama (OOD model)	74.3%	73.8	77.2	70.4

The model generalizes well to OOD splits: accuracy on emails (testenron) and unseen AI models (Llama 3.3-70B / testllama) is on par with or above the in-distribution test set.

Training Details

—Backbone: TrustSafeAI/RADAR-Vicuna-7B (frozen, 355M params)
—Dataset: pangram/editlens_iclr train split
—Examples: 1,000 per class (2,000 total), stratified sample
—Classes: human (humanwritten), `ai` (aiedited + ai_generated)
—Embedding dim: 1024
—Prototype weight: 0.3, Neural weight: 0.7
—Training time: ~6 minutes on CPU

Live Predictions Dataset

Predictions made through the hosted Space are continuously logged to adaptive-classifier/ai-detector-data — a public dataset of real-world predictions with optional user feedback (Correct / Incorrect). This dataset grows over time and can be used to track model performance, find failure cases, and drive future retraining.

Limitations

—Binary only (human vs AI) — does not distinguish AI-edited from AI-generated
—Relies on frozen RADAR embeddings; cannot learn new text patterns beyond what RADAR captures
—Minimum ~50 words of text recommended for reliable detection
—Trained on English text from specific domains (reviews, news, creative writing, academic)

Citation

bibtex

@software{adaptive_classifier,
  title = {Adaptive Classifier: Dynamic Text Classification with Continuous Learning},
  author = {Sharma, Asankhaya},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/codelion/adaptive-classifier}
}