Update model card with feedback_only retrain metrics

README.md

@@ -1,74 +1,121 @@
 ---
-language: multilingual
+language: en
 tags:
 - adaptive-classifier
 - text-classification
+- ai-detection
+- ai-generated-text
 - continuous-learning
 license: apache-2.0
+datasets:
+- pangram/editlens_iclr
+- adaptive-classifier/ai-detector-data
+base_model: TrustSafeAI/RADAR-Vicuna-7B
+metrics:
+- accuracy
+- f1
+pipeline_tag: text-classification
+model-index:
+- name: adaptive-classifier/ai-detector
+  results:
+  - task:
+      type: text-classification
+      name: AI Text Detection (Binary)
+    dataset:
+      name: EditLens ICLR 2026
+      type: pangram/editlens_iclr
+      split: test
+    metrics:
+    - type: accuracy
+      value: 74.2
+      name: Accuracy
+    - type: f1
+      value: 73.7
+      name: Macro F1
 ---
 
-# Adaptive Classifier
+# AI Text Detector (adaptive-classifier)
 
-This model is an instance of an [adaptive-classifier](https://github.com/codelion/adaptive-classifier) that allows for continuous learning and dynamic class addition.
+A binary AI text detector that classifies text as **human-written** or **AI-generated/edited**, built with [adaptive-classifier](https://github.com/codelion/adaptive-classifier) on the [EditLens ICLR 2026](https://huggingface.co/datasets/pangram/editlens_iclr) benchmark.
 
-## Installation
+## How It Works
 
-**IMPORTANT:** To use this model, you must first install the `adaptive-classifier` library. You do **NOT** need `trust_remote_code=True`.
+Uses frozen embeddings from [TrustSafeAI/RADAR-Vicuna-7B](https://huggingface.co/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
 ```
 
-## Model Details
+## Usage
+
+```python
+from adaptive_classifier import AdaptiveClassifier
+
+classifier = AdaptiveClassifier.from_pretrained("adaptive-classifier/ai-detector")
 
-- Base Model: TrustSafeAI/RADAR-Vicuna-7B
-- Number of Classes: 2
-- Total Examples: 400
-- Embedding Dimension: 1024
+predictions = classifier.predict("Your text here")
+# Returns: [('ai', 0.85), ('human', 0.15)]
 
-## Class Distribution
+# Batch prediction
+results = classifier.predict_batch(["text 1", "text 2"], k=2)
 
-```
-ai: 200 examples (50.0%)
-human: 200 examples (50.0%)
+# Continuous learning — add new examples without retraining
+classifier.add_examples(
+    ["new human text example", "new ai text example"],
+    ["human", "ai"]
+)
 ```
 
-## Usage
+## Results
 
-After installing the `adaptive-classifier` library, you can load and use this model:
+Evaluated on the [EditLens ICLR 2026](https://huggingface.co/datasets/pangram/editlens_iclr) test splits.
 
-```python
-from adaptive_classifier import AdaptiveClassifier
+### Binary Classification (Human vs AI)
 
-# Load the model (no trust_remote_code needed!)
-classifier = AdaptiveClassifier.from_pretrained("adaptive-classifier/model-name")
+| 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** |
 
-# Make predictions
-text = "Your text here"
-predictions = classifier.predict(text)
-print(predictions)  # List of (label, confidence) tuples
+### Per-Split Results
 
-# Add new examples for continuous learning
-texts = ["Example 1", "Example 2"]
-labels = ["class1", "class2"]
-classifier.add_examples(texts, labels)
-```
+| 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 |
 
-**Note:** This model uses the `adaptive-classifier` library distributed via PyPI. You do **NOT** need to set `trust_remote_code=True` - just install the library first.
+The model generalizes well to OOD splits: accuracy on emails (test_enron) and unseen AI models (Llama 3.3-70B / test_llama) is on par with or above the in-distribution test set.
 
 ## Training Details
 
-- Training Steps: 6
-- Examples per Class: See distribution above
-- Prototype Memory: Active
-- Neural Adaptation: Active
+- **Backbone**: [TrustSafeAI/RADAR-Vicuna-7B](https://huggingface.co/TrustSafeAI/RADAR-Vicuna-7B) (frozen, 355M params)
+- **Dataset**: [pangram/editlens_iclr](https://huggingface.co/datasets/pangram/editlens_iclr) train split
+- **Examples**: 1,000 per class (2,000 total), stratified sample
+- **Classes**: `human` (human_written), `ai` (ai_edited + 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](https://huggingface.co/spaces/adaptive-classifier/ai-detector) are continuously logged to [adaptive-classifier/ai-detector-data](https://huggingface.co/datasets/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
 
-This model:
-- Requires at least 3 examples per class
-- Has a maximum of 1000 examples per class
-- Updates prototypes every 100 examples
+- 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