v0.2.0

New Scorers

• bert, lora, peft - by @voorhs, @SeBorgey, @nikiduki, @riapush
• rnn, cnn - by @voorhs, @SeBorgey
• catboost - by @Samoed, @nikiduki
• zero shot methods (bi encoder, cross encoder, llm) - by @voorhs, @Darinochka

AutoML

• optimization presets by @voorhs
• refactor HPO schema by @voorhs
• autointent interruption handling by @Samoed

Other

• node validation by @Samoed
• codecarbon callback by @Darinochka, @Samoed

Chores

• updated docs and tutorials by @voorhs
• optimized tests for repo by @Samoed
• innumerable bugs fixes by @voorhs, @Samoed

Full Changelog: v0.1.0...v0.2.0

v0.1.0

New functionality

• optuna samplers: TPE, Random, Brute
• cross-validation (previously: only hold-out validation)
• basic presets for balancing between the quickness and the quality
• logging to wandb and tensorboard
• LLM-based augmentation strategies for enriching your training data

Improvements

• better regular expressions support
• better UX on conducting experiments
• more convenient way to dump fitted pipeline to disk and then load it for inference

Documentation

Check out our updated user guides!

v0.0.1

Features

• Library of intent classification methods:
  • regexp module for rule-based classification
  • proxy tuning hyperparams of embedding model using retrieval metrics
  • scoring modules for predicting intents probabilities
  • decision-making modules for constructing final prediction for multi-class and multi-label classification and out-of-domain detection
• Auto ML approach to creating intent classifier:
  • greedy optimization for tuning hyperparameters
  • no target leakage, thanks to hold-out validation
  • embeddings caching for improving efficiency
  • both Python API and CLI
• Basic but flexible inference with automatically configured intent classifier
• Easy data manipulation with hugging face datasets integrated

Documentation

• API Reference for all modules, metrics and etc
• User guides with basic and advanced usage both for Python API and CLI
• Theoretical sections on dialogue systems creation and key concepts of AutoIntent