v1.0.0 - Initial Release

LabelFusion v1.0.0

This is the first stable release of LabelFusion, introducing a unified framework for fusing LLMs and Transformer-based classifiers to achieve robust text classification performance.

Core Features

Fusion Ensemble (Core Innovation)

• AutoFusionClassifier: One-line interface for combining ML and LLM predictions.
• FusionMLP: Trainable neural fusion layer for optimal prediction weighting.
• Smart Training: Separate, adaptive learning rates for the ML backbone and fusion layer.
• Calibration Tools: Temperature scaling and isotonic regression for improved probability estimates.
• Production-Ready Utilities: Built-in caching, structured result logging, and LLM cost tracking.

Supported Models

• LLM Providers: OpenAI GPT, Google Gemini, DeepSeek.
• ML Models: Fine-tuned RoBERTa-based classifiers.
• Traditional Ensembles: Voting, weighted fusion, and class-specific routing strategies.

Classification Support

• Multi-class Classification: Single-label setups for tasks with mutually exclusive categories.
• Multi-label Classification: Fully supported, including 28-label emotion classification (e.g., GoEmotions).

Production Features

• LLM Response Caching: Automatic disk-based caching to minimize redundant API calls and reduce cost.
• Results Management: Experiment tracking, metric storage, and prediction management.
• Batch Processing: Efficient handling of large datasets.
• Async Support: Asynchronous LLM API calls for improved throughput.

Additional Notes

• Includes tests for fusion logic, calibration, and model integration.
• Provides examples and documentation for end-to-end workflows.
• First public release aligned with the JOSS submission.