facebookresearch/ParlAIArchived

Features

• Conversational AI Frameworks - Provides a comprehensive framework for building, training, and evaluating interactive conversational AI models.
• Conversational Model Training - Provides a comprehensive platform for training dialogue agents on specified datasets with configurable hyperparameters.
• Sparse-Dense Hybrid Retrievers - Combines sparse TF-IDF and dense FAISS indexes to ground conversational responses in external knowledge sources.
• Conversational Response Generation - Produces AI-driven conversational responses using sequence-to-sequence generators.
• Long-term Memory Stores - Implements persistent storage mechanisms to retain personal knowledge and conversation history across multiple dialogue turns.
• Response Grounding - Synthesizes responses grounded in verified facts retrieved from external knowledge bases.
• Hybrid Retrieval Engines - Combines FAISS and TF-IDF indexing to ground conversational responses in external information sources.
• Agent Class Abstractions - Enables the creation of custom models by inheriting base agent classes and defining specific training and evaluation logic.
• Pretrained Agent Execution - Enables the execution of saved agent models within target environments to observe and evaluate behavioral performance.
• Streaming Dataset Loaders - Streams large dialogue datasets from disk in chunks to maintain memory efficiency during training.
• Generative Model Training Tools - Provides a command-line interface for training and fine-tuning generative models on custom datasets.
• Pre-trained Model Zoos - Maintains a centralized repository of pretrained neural network architectures ready for deployment or fine-tuning.
• Model Registries - Provides a centralized registry for distributing pretrained models with versioned checkpoints and metadata.
• Dialogue Task Abstractions - Defines dialogue tasks as teacher agents that provide observations and labels through a standardized interface.
• Pretrained Model Zoos - Deno AI provides programmatic access to download and load pretrained dialogue models from a shared repository.
• Multi-Task Samplings - Supports training and evaluation across multiple datasets by sampling tasks according to configurable weights.
• PyTorch Training Loops - Implements PyTorch-based training loops for conversational agents with support for multi-task weighted sampling.
• Model Evaluation and Benchmarking - Offers integrated tools for measuring agent performance through automated benchmarks and human-in-the-loop evaluation workflows.
• Model Fine-Tuning - Optimizes pretrained transformer models for new dialogue tasks by initializing from stored zoo checkpoints.
• Dialogue Data Collectors - Provides a comprehensive system for gathering human-AI interaction data via crowdsourcing and interactive chat interfaces.
• Unified Data Access Interfaces - Provides a unified interface to load and stream diverse conversational datasets across different sources.
• Custom Teacher Agents - Provides an abstraction for teacher agents that format and serve dataset examples to student models.
• Dialogue Episode Loops - Manages multi-turn conversation flows using a world object to control turn order and termination.
• Persona-Conditioned Generation - Generates dialogue responses conditioned on specific persona descriptions to maintain character consistency.
• Conversational Model Benchmarking - Measures the quality of conversational models against specific validation datasets to assess performance.
• Conversational AI Frameworks - Serves as a research framework for training, evaluating, and sharing neural dialogue models using a unified interface.
• AI Safety Guardrails - Detects toxic or unsafe content in single-turn and multi-turn conversations using pretrained safety classifiers.
• Crowdsourcing Submission Approvers - Programmatically approves or rejects completed crowdsourcing tasks based on specific completion criteria.
• Baseline Model Architectures - Runs pre-built dialogue models or human baselines to establish performance benchmarks for comparison.
• Batched Response Generation - Produces multiple model completions for fixed utterances by pairing models with datasets and logging the results.
• Command-Line - Provides a terminal-based conversational interface for interacting with trained AI models in real time.
• Retrieval-Augmented Generation - Combines a DPR retriever with a BART generator to produce responses grounded in Wikipedia passages.
• Crowdsourced Dialogue Collection - Collects natural language dialogue data from human workers through custom tasks deployed on external platforms.
• Pretrained Model Snapshots - Maintains a centralized model zoo providing versioned pretrained checkpoints for immediate deployment and benchmarking.
• Automated Dataset Evaluation - Computes standard performance metrics on a held-out dataset to measure model quality after training.
• Dialogue Speaker Identification - Predicts which character spoke a given utterance based on conversation history and character profiles.
• Dynamic Dialogue Logic - Enables the creation of custom agents that adjust responses based on real-time input instead of fixed logs.
• Encoder-Decoder Architectures - Implements neural network designs that map input sequences to output sequences via intermediate representations using LSTM-based encoders and decoders.
• External LLM API Wrappers - Integrates external GPT-3 APIs as wrapper agents for dialogue reply generation.
• External Language Model Wrappers - Wraps external language model APIs as drop-in agents to integrate them into the research framework.
• Crowdsourced Model Evaluators - Integrates dialogue models with external crowdsourcing platforms to collect human judgments.
• Human Agent Connectors - Connects human agents to dialogue worlds via messaging platforms to facilitate data collection.
• Human-Human Dialogue Collectors - Sets up multi-turn conversations between two human participants to collect natural dialogue data.
• Image-Grounded Dialogue Generators - Generates conversational responses conditioned on image inputs for multimodal dialogue.
• Interactive AI Conversations - Gathers dialogue data from human participants interacting with a model during a task.
• Agentic Interaction Patterns - Defines interaction loops and environments to manage how multiple agents exchange messages in sequences or batches.
• Beam Search Implementations - Implements beam search decoding with configurable beam size and n-gram blocking for sequence generation.
• Model Bias Mitigation - Implements specialized training techniques to prevent dialogue models from defaulting to societal or gender biases.
• Autoregressive Model Interfaces - Provides a standard interface for autoregressive models to automate their training and evaluation processes.
• Dialogue Consistency Optimization - Uses specialized training methods to reduce contradictions and improve the coherence of chatbot responses.
• Model Self-Chat Generation - Simulates conversations between two model instances and logs the dialogue for analysis.
• Persona-Conditioned Dialogue - Loads and executes dialogue datasets where conversations are conditioned on specific user profiles or personas.
• Word Embeddings - Incorporates pre-trained vectors from Fasttext or GLOVe to initialize model embeddings for improved linguistic representation.
• Response Ranking Logic - Ranks candidate replies to select the most appropriate response for a given conversational context.
• Conversation Comparators - Facilitates side-by-side comparison of two full conversations for human workers to select the superior response.
• Task-Oriented Dialogue Simulation - Simulates goal-directed conversations by interacting with API schemas to resolve specific user requests.
• Training and Testing Splits - Implements dataset partitioning into training, validation, and testing sets via file path suffixes.
• Vocabulary Generators - Generates token-to-index mappings from task text data for model vectorization.
• Agent Behavior Simulation - Simulates conversations between agents using specific models and personas to evaluate their conversational behavior.
• Human-Model Chat Evaluators - Facilitates interactive sessions where humans converse with model agents and annotate responses.
• Multimodal Dialogue and Interaction - Supports the development of conversational agents capable of processing and generating responses grounded in visual content.
• Model Card Generation - Produces structured documentation describing a model's intended use, training data, and evaluation results via standardized model cards.
• Conversation Annotators - Presents pre-recorded conversations to humans for annotating speaker responses using checkboxes.
• Dataset Loading - Provides specialized mechanisms for loading conversational datasets from JSON files into the research framework.
• Crowdsourcing Task Builders - Enables the creation of custom dialogue logic and worker onboarding by subclassing base world and blueprint classes.
• Conversational Task Definitions - Defines new conversational tasks by implementing dedicated data building scripts and teacher agents.
• Worker Qualification Management - Filters crowdsourced participants through onboarding stages and qualification checks to ensure data quality.
• Crowdsourcing Task Blueprints - Implements a YAML-configurable overworld-subworld pattern for defining multi-stage human evaluation tasks.
• Interactive Model Interfaces - Provides a live web-based interface to send messages to trained models and inspect generated responses and metadata.
• Generative Models - Platform for training and evaluating dialogue research models.

Star history