SDialog is a modular, extensible Python toolkit for synthetic dialogue generation and analysis, designed for research and development with instruction-tuned Large Language Models (LLMs). It enables flexible, persona-driven, multi-agent dialogue simulation, orchestration, and scenario management, making it ideal for building, evaluating, and experimenting with conversational agents.
Modern conversational AI research and applications increasingly require high-quality, flexible, and reproducible synthetic dialogues for training, evaluation, and benchmarking. SDialog addresses the need for:
- Standardization: Clear definitions for dialogue, persona, and event structures.
- Abstraction: Abstract interfaces for both single-agent and multi-agent dialogue generation.
- Fine-grained Control: Orchestration to inject instructions, simulate user behaviors, and enforce scenario constraints.
- LLM Integration: Seamless integration with instruction-tuned LLMs, prompt management, and memory handling.
- Scenario and Dataset Management: Tools for managing complex scenarios, flowcharts, and persona definitions.
- Persona-based Role-Playing: Define rich agent personas to simulate realistic conversations.
- Multi-Agent Dialogue: Generate dialogues between multiple agents, each with their own persona and behavior.
- Dialogue Orchestration: Control agent actions and inject instructions dynamically using orchestrators.
- Scenario Management: Easily describe and manage dialogue scenarios, including flowcharts and user/system goals.
- Flexible Serialization: Export dialogues and events in JSON or plain text for downstream tasks.
- Integration with LLMs: Out-of-the-box support for Ollama and LangChain, with planned support for HuggingFace models.
pip install sdialogNote: You must have Ollama running on your system to use the default LLM integration.
curl -fsSL https://ollama.com/install.sh | sh
Define personas, create agents, and generate a dialogue:
from sdialog import Persona, PersonaAgent
# Define personas
alice = Persona(name="Alice", role="friendly barista", personality="cheerful and helpful")
bob = Persona(name="Bob", role="customer", personality="curious and polite")
# Create agents
alice_agent = PersonaAgent("llama2", persona=alice, name="Alice")
bob_agent = PersonaAgent("llama2", persona=bob, name="Bob")
# Generate a dialogue
dialog = alice_agent.dialog_with(bob_agent)
dialog.print()Add orchestration to control dialogue length or simulate agent behaviors:
from sdialog.orchestrators import LengthOrchestrator, ChangeMindOrchestrator
length_orch = LengthOrchestrator(min=3, max=6)
mind_orch = ChangeMindOrchestrator(probability=0.5, reasons=["changed plans", "new information"], max_times=1)
alice_agent = alice_agent | length_orch | mind_orchWork with the STAR dataset for scenario-driven dialogue generation:
from sdialog.datasets import STAR
STAR.set_path("/path/to/star-dataset")
scenario = {
"Domains": ["banking"],
"UserTask": "Open a new account",
"WizardTask": "Assist with account opening",
"Happy": True,
"MultiTask": False,
"WizardCapabilities": [{"Task": "open_account", "Domain": "banking"}]
}
system_agent, user_agent = STAR.get_agents_for_scenario(scenario, "llama2")
dialog = system_agent.dialog_with(user_agent)
dialog.print()- Documentation - Full package documentation, including installation, API reference, usage guides, and advanced examples available.
- API Reference: See docstrings in the codebase for detailed documentation of all classes and functions.
- Tutorials: Tutorials for hands-on examples as Jupyter Notebooks.
This work was supported by the EU Horizon 2020 project ELOQUENCE (grant number 101070558).
This work was also initially created in preparation for the 2025 Jelinek Memorial Summer Workshop on Speech and Language Technologies (JSALT 2025) as part of the work done by the "Play your Part" research group.
MIT License
Copyright (c) 2025 Idiap Research Institute