CoALM-8B: Conversational Agentic Language Model

Model Description

Model Sources

• 📝 Paper: https://arxiv.org/abs/2502.08820
• 🌐 Project Page: https://emrecanacikgoz.github.io/CoALM/
• 💻 Repository: https://github.com/oumi-ai/oumi/tree/main/configs/projects/calm
• 💎 Dataset: https://huggingface.co/datasets/uiuc-convai/CoALM-IT

Model Details

• Model Name: CoALM-8B
• Developed by: Colloboration of UIUC Conversational AI LAB and Oumi
• License: cc-by-nc-4.0
• Architecture: Fine-tuned Llama 3.1 8B Instruct
• Training Data: CoALM-IT dataset
• Fine-tuning Framework: Oumi
• Training Hardware: 8 NVIDIA H100 GPUs
• Training Duration: ~8 hours
• Evaluation Benchmarks: MultiWOZ 2.4, BFCL V3, API-Bank
• Release Date: February 5, 2025

Capabilities and Features

🗣 Conversational Agentic Abilities

• Multi-turn Dialogue Mastery: Maintains coherent conversations across multiple turns with accurate state tracking.
• Function Calling and API Integration: Dynamically selects and calls APIs for task execution.
• ReAct-based Reasoning: Utilizes a structured reasoning process (User-Thought-Action-Observation-Thought-Response).
• Zero-Shot Generalization: Excels in previously unseen function-calling tasks.

🚀 Benchmark Performance

• MultiWOZ 2.4 (TOD): Excels in dialogue state tracking and task completion.
• BFCL V3 (LA): Demonstrates superior function-calling abilities over language agents.
• API-Bank (LA): Accurately generates API calls and integrates responses into conversation flow.

Training Process

🔧 Fine-tuning Stages

1. TOD Fine-tuning: Optimized for dialogue state tracking (e.g., augmented SNIPS reformatted in Alpaca-style instruction tuning).
2. Function Calling Fine-tuning: Trained to select and generate well-formed API calls from LA datasets.
3. ReAct-based Fine-tuning: Addresses multi-turn conversations with API integration using a structured reasoning framework.

🔍 Training Hyperparameters

• Base Model: Llama 3.1 8B Instruct
• LoRA Config: Rank = 16, Scaling Factor = 32
• Batch Size: 8
• Learning Rate: 1e-4
• Optimizer: AdamW (betas = 0.9, 0.999, epsilon = 1e-8)
• Precision: Mixed precision (bfloat16)
• Warm-up Steps: 0.1 ratio of total steps
• Gradient Accumulation Steps: 1

💡 CoALM-IT Dataset

📊 Benchmark Performance

Usage

🏗 How to Load the Model using Transformers

🛠 Example Oumi Inference

🛠 Example Oumi Fine-Tuning

• Task-Specific Calibration: While CoALM-8B generalizes well across tasks, performance can improve with domain-specific fine-tuning.
• Scalability to Larger Models: Future iterations (CoALM-70B, CoALM-405B) extend capabilities to larger-scale agentic conversations.
• Open-Source Expansion: All datasets, training scripts, and model checkpoints are publicly available to foster further research.

Acknowledgements

License

Citation

@misc{acikgoz2025singlemodelmastermultiturn,
      title={Can a Single Model Master Both Multi-turn Conversations and Tool Use? CoALM: A Unified Conversational Agentic Language Model}, 
      author={Emre Can Acikgoz and Jeremiah Greer and Akul Datta and Ze Yang and William Zeng and Oussama Elachqar and Emmanouil Koukoumidis and Dilek Hakkani-Tür and Gokhan Tur},
      year={2025},
      eprint={2502.08820},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2502.08820}, 
}