Adaptive Classifier

🚀 Dynamic text classification with continuous learning, strategic defense, and zero-downtime adaptation

---

🔗 Quick Links

• 📚 HuggingFace Organization - Pre-trained models and datasets
• 📖 Articles & Tutorials:
  • Adaptive Classifier: Dynamic Text Classification
  • AutoThink: Advanced Reasoning Techniques
  • Enterprise-Ready Classifiers

---

Adaptive Classifier is a PyTorch-based machine learning library that revolutionizes text classification with continuous learning, dynamic class addition, and strategic defense against adversarial inputs. Built on HuggingFace transformers, it enables zero-downtime model updates and enterprise-grade robustness.

✨ Key Features

🎯 Core Capabilities

• 🚀 Universal Compatibility - Works with any HuggingFace transformer model
• ⚡ Optimized Inference - Built-in ONNX Runtime for 2-4x faster CPU predictions
• 📈 Continuous Learning - Add new examples without catastrophic forgetting
• 🔄 Dynamic Classes - Add new classes at runtime without retraining
• ⏱️ Zero Downtime - Update models in production without service interruption
• ✅ Knows What It Doesn't Know - Calibrated confidence, prediction sets, out-of-distribution detection and abstention
• 🧹 Fixable - Forget a wrong example or a whole class without retraining
• 🎣 Active Learning & Drift - Suggest which texts to label next and detect when traffic has shifted
• 🔌 Easy to Adopt - scikit-learn interface, HTTP server, async methods, and ONNX inference outside Python

🛡️ Advanced Defense

• 🎮 Strategic Classification - Game-theoretic defense against adversarial manipulation
• 🔒 Anti-Gaming Protection - Robust predictions under strategic behavior
• ⚖️ Multiple Prediction Modes - Regular, strategic, and robust inference options

🧠 Intelligent Architecture

• 💾 Prototype Memory - FAISS-powered efficient similarity search
• 🔬 Adaptive Neural Layer - Trainable classification head with EWC protection
• 🎯 Hybrid Predictions - Combines prototype similarity and neural network outputs
• 📊 HuggingFace Integration - Push/pull models directly from the Hub

---

📊 Performance & Benchmarks

🛡️ Strategic Classification Defense

Tested on adversarial examples from AI-Secure/adv_glue dataset:

Metric	Regular Classifier	Strategic Classifier	Improvement
Clean Data Accuracy	80.00%	82.22%	+2.22%
Adversarial Data Accuracy	60.00%	82.22%	+22.22%
Robustness (vs attack)	-20.00% drop	0.00% drop	Perfect

🔍 Hallucination Detection

Evaluated on RAGTruth benchmark across multiple task types:

Task Type	Precision	Recall	F1 Score
QA	35.50%	45.11%	39.74%
Summarization	22.18%	96.91%	36.09%
Data-to-Text	65.00%	100.0%	78.79%
Overall	40.89%	80.68%	51.54%

🚦 LLM Router Optimization

Tested on arena-hard-auto-v0.1 dataset (500 queries):

Metric	Without Adaptation	With Adaptation	Improvement
Cost Savings	25.60%	32.40%	+6.80%
Efficiency Ratio	1.00x	1.27x	+27%
Resource Utilization	Standard	Optimized	Better

Key Insight: Adaptive classification maintains quality while significantly improving cost efficiency and robustness across all tested scenarios.

---

Try Now

Use Case	Demonstrates	Link
Basic Example (Cat or Dog)	Continuous learning
Support Ticket Classification	Realistic examples
Query Classification	Different configurations
Multilingual Sentiment Analysis	Ensemble of classifiers
Product Category Classification	Batch processing
Multi-label Classification	Extensibility

🚀 Installation

Quick Install

pip install adaptive-classifier

Includes: ONNX Runtime for 2-4x faster CPU inference out-of-the-box

🛠️ Development Setup

# Clone the repository
git clone https://github.com/codelion/adaptive-classifier.git
cd adaptive-classifier

# Install in development mode with test dependencies
pip install -e ".[test]"

---

Pooling

AdaptiveClassifier reduces a model's token embeddings to one vector per text. How it does that matters, because prototype memory is cosine nearest-neighbour search over those vectors.

Before 0.2.0 the CLS token was always used. That is wrong for most encoders. all-MiniLM-L6-v2 publishes pooling_mode_mean_tokens, so it is a mean-pooling model, and a masked language model never trains its CLS token as a sentence representation at all. On a semantic-similarity check with all-MiniLM-L6-v2, CLS scored unrelated sentence pairs at 0.53 cosine similarity while mean pooling scored them at -0.03: a 2.4x difference in separation between related and unrelated text.

From 0.2.0 the default is auto. The model's own sentence-transformers pooling config is used when it publishes one, otherwise mean pooling.

# use what the model was trained with (default)
classifier = AdaptiveClassifier("sentence-transformers/all-MiniLM-L6-v2")

# or choose explicitly
classifier = AdaptiveClassifier("bert-base-uncased", config={"pooling": "mean"})
classifier = AdaptiveClassifier("bert-base-uncased", config={"pooling": "cls"})

Classifiers saved before 0.2.0 keep CLS pooling when reloaded, so upgrading does not change their predictions. Retrain, or pass pooling explicitly, to pick up the new behaviour.

How a prediction is made

A prediction mixes two signals: the prototype score (how close the text is to each class's mean embedding) and the neural head (a small trained network on the same embeddings). prototype_weight and neural_weight (default 0.7 / 0.3) control the mix.

classifier = AdaptiveClassifier(
    "sentence-transformers/all-MiniLM-L6-v2",
    config={"prototype_weight": 0.3, "neural_weight": 0.7},
)

The best split depends on how well your encoder separates your classes: an encoder that bunches everything together produces weak prototypes and is better off leaning on the head.

Classes with few examples. Until a class has new_class_example_threshold examples (default 10), the head's weight ramps up linearly from zero to neural_weight and the prototype takes the rest, so a brand-new class is judged almost entirely by its prototype. (Before 0.3.0 such classes used a fixed 0.3 / 0.7 split in favour of the head, and a head fitted to a handful of points could confidently outvote a correct prototype. To get a fixed split back, set both new_class_prototype_weight and new_class_neural_weight.)

Recognising a new class sooner. The ramp is the dial between "protect the classes you already have" and "pick up a new class quickly". Lowering new_class_example_threshold gives the head a say earlier. Measured with MiniLM, adding classes one at a time with 4 examples each (3 seeds, 600 test texts; averages over the added classes):

new_class_example_threshold	ag_news: final accuracy / new-class recall / old-class loss	emotion: final accuracy / new-class recall / old-class loss
10 (default)	0.76 / 0.77 / 0.06	0.35 / 0.43 / 0.09
5	0.74 / 0.85 / 0.12	0.35 / 0.59 / 0.14
3	0.72 / 0.87 / 0.15	0.34 / 0.59 / 0.15
fixed 0.3 / 0.7 (pre-0.3.0 split)	0.66 / 0.91 / 0.22	0.32 / 0.68 / 0.18

Each step buys new-class recall with old-class accuracy. If a new class that is missed matters more to you than one that disturbs the others, try 5; the default favours stability. Two datasets and one encoder, so check it on your own data.

Prototype sharpness. Prototype scores are softmax(-distance² / prototype_temperature) with a default temperature of 0.25. Lower values make the nearest class stand out more. Set prototype_temperature to None for the scoring used before 0.3.0, which barely separated the nearest class from the rest.

Head training. The head is trained for at least head_steps optimiser steps (default 300) with a cosine-decayed head_learning_rate (default 0.003). That takes roughly 1-2 seconds on a small memory; lower head_steps (for example to 100) if you add examples very frequently. Before 0.3.0 the head got about ten steps, too few to learn even a simple XOR pattern, which prototypes alone cannot separate.

Classifiers saved before 0.3.0 keep their old scoring when reloaded, so upgrading does not change their predictions. Retrain, or set the keys above, to adopt the new behaviour.

⚡ Quick Start

30-Second Setup

Get started with adaptive classification in under 30 seconds:

from adaptive_classifier import AdaptiveClassifier

# 🎯 Step 1: Initialize with any HuggingFace model
classifier = AdaptiveClassifier("bert-base-uncased")

# 📝 Step 2: Add training examples
texts = ["The product works great!", "Terrible experience", "Neutral about this purchase"]
labels = ["positive", "negative", "neutral"]
classifier.add_examples(texts, labels)

# 🔮 Step 3: Make predictions
predictions = classifier.predict("This is amazing!")
print(predictions)  
# Output: [('positive', 0.85), ('neutral', 0.12), ('negative', 0.03)]

🏷️ Multi-Label Classification

Classify texts into multiple categories simultaneously with automatic threshold adaptation:

from adaptive_classifier import MultiLabelAdaptiveClassifier

# Initialize multi-label classifier
classifier = MultiLabelAdaptiveClassifier(
    "bert-base-uncased",
    min_predictions=1,    # Ensure at least 1 prediction
    max_predictions=5     # Limit to top 5 predictions
)

# Multi-label training data (each text can have multiple labels)
texts = [
    "AI researchers study climate change using machine learning",
    "Tech startup develops healthcare solutions"
]
labels = [
    ["technology", "science", "climate", "ai"],
    ["technology", "business", "healthcare"]
]

classifier.add_examples(texts, labels)

# Make multi-label predictions
predictions = classifier.predict_multilabel("Medical AI breakthrough announced")
# Output: [('healthcare', 0.72), ('technology', 0.68), ('ai', 0.45)]

💾 Save & Load Models

# Save locally
classifier.save("./my_classifier")
loaded_classifier = AdaptiveClassifier.load("./my_classifier")

# 🤗 HuggingFace Hub Integration
classifier.push_to_hub("adaptive-classifier/my-model")
hub_classifier = AdaptiveClassifier.from_pretrained("adaptive-classifier/my-model")

🎮 Strategic Defense (Anti-Gaming)

# Enable strategic classification for adversarial robustness
config = {'enable_strategic_mode': True}
strategic_classifier = AdaptiveClassifier("bert-base-uncased", config=config)

# Robust predictions against manipulation
predictions = strategic_classifier.predict("This product has amazing quality features!")
# Returns predictions that consider potential gaming attempts

⚡ Optimized CPU Inference with ONNX

Adaptive Classifier includes built-in ONNX Runtime support for 2-4x faster CPU inference with zero code changes required.

Automatic Optimization (Default)

ONNX Runtime is automatically used on CPU for optimal performance:

# Automatically uses ONNX on CPU, PyTorch on GPU
classifier = AdaptiveClassifier("bert-base-uncased")

# That's it! Predictions are 2-4x faster on CPU
predictions = classifier.predict("Fast inference!")

Performance Comparison

Configuration	Speed	Use Case
PyTorch (GPU)	Fastest	GPU servers
ONNX (CPU)	2-4x faster	Production CPU deployments
PyTorch (CPU)	Baseline	Development, training

Save & Deploy with ONNX

# Save with ONNX export (both quantized & unquantized versions)
classifier.save("./model")

# Push to Hub with ONNX (both versions included by default)
classifier.push_to_hub("username/model")

# Load automatically uses quantized ONNX on CPU (fastest, 4x smaller)
fast_classifier = AdaptiveClassifier.load("./model")

# Choose unquantized ONNX for maximum accuracy
accurate_classifier = AdaptiveClassifier.load("./model", prefer_quantized=False)

# Force PyTorch (no ONNX)
pytorch_classifier = AdaptiveClassifier.load("./model", use_onnx=False)

# Opt-out of ONNX export when saving
classifier.save("./model", include_onnx=False)

ONNX Model Versions:

• Quantized (default): INT8 quantized, 4x smaller, ~1.14x faster on ARM, 2-4x faster on x86
• Unquantized: Full precision, maximum accuracy, larger file size

By default, models are saved with both versions, and the quantized version is automatically loaded for best performance. Use prefer_quantized=False if you need maximum accuracy.

Benchmark Your Model

# Compare PyTorch vs ONNX performance
python scripts/benchmark_onnx.py --model bert-base-uncased --runs 100

Example Results:

Model: bert-base-uncased (CPU)
PyTorch:  8.3ms/query  (baseline)
ONNX:     2.1ms/query  (4.0x faster) ✓

Note: ONNX optimization is included by default. For GPU inference, PyTorch is automatically used for best performance.

Multilingual models and other runtimes (.NET, JavaScript, ...)

Any multilingual encoder works as the base model, and a saved classifier can be run from C#, Node or the browser through an ONNX runtime. See the deployment guide and the dependency-light reference implementation in examples/portable_inference.py.

Advanced Usage

scikit-learn Interface

SklearnAdaptiveClassifier works with Pipeline, cross_val_score and GridSearchCV. X is a list of strings.

from adaptive_classifier import SklearnAdaptiveClassifier
from sklearn.model_selection import cross_val_score

clf = SklearnAdaptiveClassifier("sentence-transformers/all-MiniLM-L6-v2")
clf.fit(texts, labels)
clf.predict(["Where is my refund?"])          # class labels
clf.predict_proba(["Where is my refund?"])    # columns follow clf.classes_
cross_val_score(clf, texts, labels, cv=5)

# The library's strength: keep learning, including brand-new classes
clf.partial_fit(["App crashes on login"], ["bug"])

clf.classifier_.save("./model")               # the underlying AdaptiveClassifier

fit starts over each time; partial_fit adds to what is already learned. A fitted estimator can be pickled, but that copies the whole encoder, so save classifier_ instead.

Calibrated Confidence and Prediction Sets

Raw scores are not probabilities: a model can say 0.9 and be right 70% of the time. Fit a calibration on labelled examples the model was not trained on:

info = classifier.calibrate(held_out_texts, held_out_labels)
info["ece_before"], info["ece_after"]       # expected calibration error, lower is better

classifier.predict(text)                    # confidences now match how often the model is right
classifier.predict(text, abstain_below=0.8) # so thresholds like this mean what they say

# A set of labels that contains the true one at least 90% of the time
classifier.predict_set(text, alpha=0.1)     # [("billing", 0.55), ("refunds", 0.31)]
classifier.calibration_report(test_texts, test_labels)   # check it on separate data

One easy input gives a one-label set; an ambiguous one gives several. Calibration is saved with the model and discarded if you add or forget a class; re-run it after substantial new data.

Active Learning and Drift

# Which of these 5,000 unlabeled texts should a person label next?
for index, score in classifier.suggest_labels(unlabeled, n=20, strategy="margin", diverse=True):
    print(unlabeled[index])

classifier.suggest_labels(unlabeled, n=20, strategy="ood")   # hunt for classes you have not defined yet

# Has this week's traffic moved away from what the model knows?
report = classifier.drift_report(recent_texts)
report["drifted"], report["ood_rate"], report["p_value"]

margin (the default) finds texts on a decision boundary; ood finds texts unlike any known class. Uncertainty sampling is not guaranteed to beat labeling at random: when classes overlap so much that the confusion is irreducible noise, labeling the confusing texts teaches the model little (it did not in a synthetic test with heavily overlapping classes). It earns its keep when ambiguity comes from missing data, and ood is the reliable way to find classes you have not defined yet.

Serving

pip install "adaptive-classifier[serve]"
python -m adaptive_classifier.serving ./my_classifier --port 8000
curl -s localhost:8000/predict -H 'content-type: application/json' -d '{"text": "where is my refund?"}'

Predictions, batches, conformal sets and out-of-distribution checks over HTTP, plus optional API-key-protected endpoints for adding examples and classes while it runs. One classifier is safe to share between threads, and await classifier.apredict(...) keeps an event loop responsive. See docs/serving.md and docker/Dockerfile.

Fixing Mistakes and Knowing When to Abstain

# Remove a wrongly labelled example, or a whole class
classifier.remove_examples(["Refund my order"], label="technical")
classifier.forget("obsolete_class")

# Detect inputs unlike anything the classifier has seen
classifier.is_ood("completely unrelated text")        # True / False
classifier.ood_score("completely unrelated text")     # ~1 or below = familiar, higher = further out

# Abstain instead of guessing (an empty list means "don't know")
classifier.predict(text, abstain_below=0.6)           # low confidence
classifier.predict(text, abstain_ood=True)            # out of distribution

Out-of-distribution scores compare a text's distance from the nearest class prototype to that class's own spread, so no per-model tuning of absolute distances is needed. The default ood_threshold of 1.25 is a starting point; check it on held-out data.

Adding New Classes Dynamically

# Add a completely new class
new_texts = [
    "Error code 404 appeared",
    "System crashed after update"
]
new_labels = ["technical"] * 2

classifier.add_examples(new_texts, new_labels)

Continuous Learning

# Add more examples to existing classes
more_examples = [
    "Best purchase ever!",
    "Highly recommend this"
]
more_labels = ["positive"] * 2

classifier.add_examples(more_examples, more_labels)

Multi-Label Classification with Advanced Configuration

from adaptive_classifier import MultiLabelAdaptiveClassifier

# Configure advanced multi-label settings
classifier = MultiLabelAdaptiveClassifier(
    "bert-base-uncased",
    default_threshold=0.5,      # Base threshold for predictions
    min_predictions=1,          # Minimum labels to return
    max_predictions=10          # Maximum labels to return
)

# Training with diverse multi-label examples
texts = [
    "Scientists develop AI for medical diagnosis and climate research",
    "Tech company launches sustainable energy and healthcare products",
    "Olympic athletes use sports science and nutrition technology"
]
labels = [
    ["science", "ai", "healthcare", "research"],
    ["technology", "business", "environment", "healthcare"],
    ["sports", "science", "health", "technology"]
]

classifier.add_examples(texts, labels)

# Advanced prediction options
predictions = classifier.predict_multilabel(
    "New research on AI applications in environmental science",
    threshold=0.3,     # Custom threshold
    max_labels=5       # Limit results
)

# Get detailed statistics
stats = classifier.get_label_statistics()
print(f"Adaptive threshold: {stats['adaptive_threshold']}")
print(f"Label-specific thresholds: {stats['label_thresholds']}")

Strategic Classification (Anti-Gaming)

# Enable strategic mode to defend against adversarial inputs
config = {
    'enable_strategic_mode': True,
    'cost_function_type': 'linear',
    # One cost per embedding dimension (384 for all-MiniLM-L6-v2, 768 for BERT-base):
    # how expensive it is for an adversary to push that dimension upward.
    'cost_coefficients': [0.3] * 768,
    'strategic_blend_regular_weight': 0.6,   # Weight for regular predictions
    'strategic_blend_strategic_weight': 0.4  # Weight for strategic predictions
}

classifier = AdaptiveClassifier("bert-base-uncased", config=config)
classifier.add_examples(texts, labels)
# If cost_coefficients has the wrong length, strategic mode is switched off and the
# reason is logged; check classifier.strategic_mode.

# Robust predictions that consider potential manipulation
text = "This product has amazing quality features!"

# Dual prediction (automatic blend of regular + strategic)
predictions = classifier.predict(text)

# Pure strategic prediction (simulates adversarial manipulation)
strategic_preds = classifier.predict_strategic(text)

# Robust prediction (assumes input may already be manipulated)
robust_preds = classifier.predict_robust(text)

print(f"Dual: {predictions}")
print(f"Strategic: {strategic_preds}")
print(f"Robust: {robust_preds}")

🏷️ Multi-Label Classification

The MultiLabelAdaptiveClassifier extends adaptive classification to handle scenarios where each text can belong to multiple categories simultaneously. It automatically handles threshold adaptation for scenarios with many labels.

Key Features

• 🎯 Automatic Threshold Adaptation: Dynamically adjusts thresholds based on the number of labels to prevent empty predictions
• 📊 Sigmoid Activation: Uses proper multi-label architecture with BCE loss instead of softmax
• ⚙️ Configurable Limits: Set minimum and maximum number of predictions per input
• 📈 Label-Specific Thresholds: Automatically adjusts thresholds based on label frequency
• 🔄 Incremental Learning: Add new labels and examples without retraining from scratch

Usage

from adaptive_classifier import MultiLabelAdaptiveClassifier

# Initialize with configuration
classifier = MultiLabelAdaptiveClassifier(
    "distilbert/distilbert-base-cased",
    default_threshold=0.5,
    min_predictions=1,
    max_predictions=5
)

# Multi-label training data
texts = [
    "Breaking: Scientists discover AI can help predict climate change patterns",
    "Tech giant announces breakthrough in quantum computing for healthcare",
    "Olympic committee adopts new sports technology for athlete performance"
]

labels = [
    ["science", "technology", "climate", "news"],
    ["technology", "healthcare", "quantum", "business"],
    ["sports", "technology", "performance", "news"]
]

# Train the classifier
classifier.add_examples(texts, labels)

# Make predictions
predictions = classifier.predict_multilabel(
    "Revolutionary medical AI system launched by tech startup"
)

# Results: [('technology', 0.85), ('healthcare', 0.72), ('business', 0.45)]

Adaptive Thresholds

The classifier automatically adjusts prediction thresholds based on the number of labels:

Number of Labels	Threshold	Benefit
2-4 labels	0.5 (default)	Standard precision
5-9 labels	0.4 (20% lower)	Balanced recall
10-19 labels	0.3 (40% lower)	Better coverage
20-29 labels	0.2 (60% lower)	Prevents empty results
30+ labels	0.1 (80% lower)	Ensures predictions

This solves the common "No labels met the threshold criteria" issue when dealing with many-label scenarios.

---

🏢 Enterprise Use Cases

🔍 Hallucination Detection

Detect when LLMs generate information not supported by provided context (51.54% F1, 80.68% recall):

detector = AdaptiveClassifier.from_pretrained("adaptive-classifier/llm-hallucination-detector")
context = "France is in Western Europe. Capital: Paris. Population: ~67 million."
response = "Paris is the capital. Population is 70 million."  # Contains hallucination

prediction = detector.predict(f"Context: {context}\nAnswer: {response}")
# Returns: [('HALLUCINATED', 0.72), ('NOT_HALLUCINATED', 0.28)]

🚦 Intelligent LLM Routing

Optimize costs by routing queries to appropriate model tiers (32.40% cost savings):

router = AdaptiveClassifier.from_pretrained("adaptive-classifier/llm-router")
query = "Write a function to calculate Fibonacci sequence"

predictions = router.predict(query)
# Returns: [('HIGH', 0.92), ('LOW', 0.08)]
# Route to GPT-4 for complex tasks, GPT-3.5 for simple ones

⚙️ Configuration Optimization

Automatically predict optimal LLM settings (temperature, top_p) for different query types:

config_optimizer = AdaptiveClassifier.from_pretrained("adaptive-classifier/llm-config-optimizer")
query = "Explain quantum physics concepts"

predictions = config_optimizer.predict(query)
# Returns: [('BALANCED', 0.85), ('CREATIVE', 0.10), ...]
# Automatically suggests temperature range: 0.6-1.0 for balanced responses

🛡️ Content Moderation

Deploy enterprise-ready classifiers for various moderation tasks:

# Available pre-trained enterprise classifiers:
classifiers = [
    "adaptive-classifier/content-moderation",      # Content safety
    "adaptive-classifier/business-sentiment",      # Business communications
    "adaptive-classifier/pii-detection",           # Privacy protection
    "adaptive-classifier/fraud-detection",         # Financial security
    "adaptive-classifier/email-priority",          # Email routing
    "adaptive-classifier/compliance-classification" # Regulatory compliance
]

# Easy deployment
moderator = AdaptiveClassifier.from_pretrained("adaptive-classifier/content-moderation")
result = moderator.predict("User generated content here...")

💡 Pro Tip: All enterprise models support continuous adaptation - add your domain-specific examples to improve performance over time.

---

Architecture Overview

The Adaptive Classifier combines four key components in a unified architecture:

1. Transformer Embeddings: Uses state-of-the-art language models for text representation

2. Prototype Memory: Maintains class prototypes for quick adaptation to new examples

3. Adaptive Neural Layer: Learns refined decision boundaries through continuous training

4. Strategic Classification: Defends against adversarial manipulation using game-theoretic principles. When strategic mode is enabled, the system:

   • Models potential strategic behavior of users trying to game the classifier
   • Uses cost functions to represent the difficulty of manipulating different features
   • Combines regular predictions with strategic-aware predictions for robustness
   • Provides multiple prediction modes: dual (blended), strategic (simulates manipulation), and robust (anti-manipulation)

Why Adaptive Classification?

Traditional classification approaches face significant limitations when dealing with evolving requirements and adversarial environments:

The Adaptive Classifier overcomes these limitations through:

• Dynamic class addition without full retraining
• Strategic robustness against adversarial manipulation
• Memory-efficient prototypes with FAISS optimization
• Zero downtime updates for production systems
• Game-theoretic defense mechanisms

Continuous Learning Process

The system evolves through distinct phases, each building upon previous knowledge without catastrophic forgetting:

The learning process includes:

• Initial Training: Bootstrap with basic classes
• Dynamic Addition: Seamlessly add new classes as they emerge
• Continuous Learning: Refine decision boundaries with EWC protection
• Strategic Enhancement: Develop robustness against manipulation
• Production Deployment: Full capability with ongoing adaptation

Order Dependency in Online Learning

When using the adaptive classifier for true online learning (adding examples incrementally), be aware that the order in which examples are added can affect predictions. This is inherent to incremental neural network training.

The Challenge

# These two scenarios may produce slightly different models:

# Scenario 1
classifier.add_examples(["fish example"], ["aquatic"])
classifier.add_examples(["bird example"], ["aerial"])

# Scenario 2  
classifier.add_examples(["bird example"], ["aerial"])
classifier.add_examples(["fish example"], ["aquatic"])

While we've implemented sorted label ID assignment to minimize this effect, the neural network component still learns incrementally, which can lead to order-dependent behavior.

Solution: Prototype-Only Predictions

For applications requiring strict order independence, you can configure the classifier to rely solely on prototype-based predictions:

# Configure to use only prototypes (order-independent)
config = {
    'prototype_weight': 1.0,  # Use only prototypes
    'neural_weight': 0.0      # Disable neural network contribution
}

classifier = AdaptiveClassifier("bert-base-uncased", config=config)

With this configuration:

• Predictions are based solely on similarity to class prototypes (mean embeddings)
• Results are completely order-independent
• Trade-off: May have slightly lower accuracy than the hybrid approach

Best Practices

1. For maximum consistency: Use prototype-only configuration
2. For maximum accuracy: Accept some order dependency with the default hybrid approach
3. For production systems: Consider batching updates and retraining periodically if strict consistency is required
4. Model selection matters: Some models (e.g., google-bert/bert-large-cased) may produce poor embeddings for single words. For better results with short inputs, consider:
   • bert-base-uncased
   • sentence-transformers/all-MiniLM-L6-v2
   • Or any model specifically trained for semantic similarity

---

🔗 Related Projects

• OpenEvolve - Open-source evolutionary coding agent for algorithm discovery
• OptiLLM - Optimizing inference proxy with 20+ techniques for 2-10x accuracy improvements

🤝 Community & Contributing

• 🐛 Issues & Bug Reports: GitHub Issues
• 💬 Discussions: GitHub Discussions
• 📖 Documentation: API Reference
• 🛠️ Contributing: CONTRIBUTING.md

📚 References

• Strategic Classification
• RouteLLM: Learning to Route LLMs with Preference Data
• Transformer^2: Self-adaptive LLMs
• Lamini Classifier Agent Toolkit
• Protoformer: Embedding Prototypes for Transformers
• Overcoming catastrophic forgetting in neural networks
• RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models
• LettuceDetect: A Hallucination Detection Framework for RAG Applications

📜 Citation

If you use this library in your research, please cite:

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

---

Made with ❤️ by Adaptive Classifier Team

⭐ Star us on GitHub • 🤗 Follow on HuggingFace