CoALM-8B is an 8 billion parameter Conversational Agentic Language Model developed by UIUC Conversational AI LAB and Oumi, fine-tuned from Llama 3.1 8B Instruct. It unifies Task-Oriented Dialogue (TOD) capabilities with Language Agent (LA) functionalities, excelling at multi-turn reasoning and complex API usage. The model is optimized for dialogue state tracking, function calling, and ReAct-based reasoning, outperforming domain-specific models on benchmarks like MultiWOZ 2.4, BFCL V3, and API-Bank.
Loading preview...
CoALM-8B: Conversational Agentic Language Model
CoALM-8B is an 8 billion parameter model developed by the UIUC Conversational AI LAB and Oumi, fine-tuned from Llama 3.1 8B Instruct. It represents the smallest open-source model in the CoALM series, designed to integrate both Task-Oriented Dialogue (TOD) and Language Agent (LA) functionalities into a unified system. The model was fine-tuned on CoALM-IT, a novel dataset that interleaves multi-turn ReAct-based reasoning with complex API usage.
Key Capabilities
- Unified TOD and LA: Seamlessly combines conversational understanding with tool-use capabilities.
- Multi-turn Dialogue Mastery: Maintains coherent conversations and accurate state tracking across multiple turns.
- Function Calling & API Integration: Dynamically selects and executes APIs for task completion.
- ReAct-based Reasoning: Employs a structured reasoning process (User-Thought-Action-Observation-Thought-Response) for complex interactions.
- Zero-Shot Generalization: Demonstrates strong performance on unseen function-calling tasks.
- Benchmark Performance: Outperforms top domain-specific models on key benchmarks including MultiWOZ 2.4 (TOD), BFCL V3 (LA), and API-Bank (LA).
Good For
- Developing advanced conversational agents that require both dialogue management and external tool use.
- Applications needing robust function-calling and API integration within multi-turn interactions.
- Research into unified models for conversational AI and language agents.
CoALM-8B's training involved distinct stages for TOD, function calling, and ReAct-based fine-tuning, utilizing the Oumi framework on 8 NVIDIA H100 GPUs. All datasets, training scripts, and model checkpoints are publicly available to foster further research.