|
--- |
|
license: cc-by-nc-4.0 |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
base_model: |
|
- meta-llama/Llama-3.1-405B-Instruct |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
# CoALM-405B: The Largest Open-Source Agentic LLM |
|
|
|
[](https://github.com/oumi-ai/oumi) |
|
|
|
## 🌟 Model Overview |
|
|
|
**CoALM-405B** is the **largest fully open-source Conversational Agentic Language Model**. This model sets a new standard in **Conversational AI**, seamlessly integrating both **Task-Oriented Dialogue (TOD) capabilities** and **Language Agent (LA) functionalities**. |
|
It is designed to **push the boundaries** of open-source agentic LLMs, excelling at **multi-turn dialogue, tool usage, reasoning, and API execution**. It is the **best-performing fully open-source LLM** on the **Berkeley Function Calling Leaderboard V3 (BFCL V3)**, marking a leap in open-source AI research. |
|
|
|
## Model Sources |
|
|
|
<!-- Provide the basic links for the model. --> |
|
|
|
- 📝 **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-405B |
|
- **Developed by:** Colloboration of UIUC Conversational AI LAB and Oumi |
|
- **License:** cc-by-nc-4.0 |
|
- **Architecture:** Meta-Llama 3.1-405B Instruct |
|
- **Training Data:** CoALM-IT |
|
- **Fine-tuning Framework:** [Oumi](https://github.com/oumi-ai/oumi) |
|
- **Training Hardware:** 8 NVIDIA H100 GPUs |
|
- **Training Duration:** ~6.5 days |
|
- **Evaluation Benchmarks:** MultiWOZ 2.4, BFCL V3, API-Bank |
|
- **Release Date:** February 5, 2025 |
|
|
|
--- |
|
## 🏆 Why CoALM-405B is a Game-Changer |
|
|
|
- **🚨 Largest Open-Source Agentic LLM:** A **405B** parameter model that brings state-of-the-art agentic capabilities to the public domain. |
|
- **🎯 Best Open-Source Performance on BFCL V3:** Outperforms leading proprietary models like **GPT-4o, Gemini, and Claude** in function-calling tasks. |
|
- **🔍 True Zero-Shot Function Calling:** Generalizes to unseen API tasks with **unmatched accuracy**. |
|
- **🤖 Multi-Turn Dialogue Mastery:** Excels at long conversations, **task tracking, and complex reasoning**. |
|
- **🛠 API Tool Use and Reasoning:** Makes precise API calls, interprets responses, and synthesizes **coherent** multi-step solutions. |
|
- **📜 Fully Open-Source & Reproducible:** Released under **cc-by-nc-4.0**, including model weights, training logs, and datasets. |
|
|
|
|
|
## 💡 CoALM-IT Dataset |
|
<img src="table.png" alt="CALM-IT Dataset Statistics" width="800"/> |
|
|
|
|
|
--- |
|
## 📊 Benchmark Performance |
|
|
|
<img src="results.png" alt="CALM-IT Dataset Statistics" width="1000"/> |
|
|
|
--- |
|
## 🔧 Training Process |
|
### Fine-tuning Stages |
|
1. **TOD Fine-tuning:** Optimized for **dialogue state tracking** (e.g., augmented SNIPS in instruction-tuned format). |
|
2. **Function Calling Fine-tuning:** Trained to generate **highly accurate API calls** from LA datasets. |
|
3. **ReAct-based Fine-tuning:** Enhances multi-turn conversations with structured **thought-action-observation-response reasoning**. |
|
|
|
### Training Hyperparameters |
|
- **Base Model:** Meta-Llama 3.1-405B Instruct |
|
- **LoRA Config:** Rank = 16, Scaling Factor = 32 |
|
- **Batch Size:** 2 |
|
- **Learning Rate:** 1e-4 |
|
- **Optimizer:** AdamW (betas = 0.9, 0.999, epsilon = 1e-8) |
|
- **Precision:** q4 |
|
- **Warm-up Steps:** 500 |
|
- **Gradient Accumulation Steps:** 1 |
|
|
|
--- |
|
|
|
## ❗️ How to Use CoALM-405B |
|
It requires 16xH100 NVIDIA GPUs for Inference. |
|
|
|
### 🏗 How to Load the Model using HuggingFace |
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
tokenizer = AutoTokenizer.from_pretrained("uiuc-convai/CoALM-8B") |
|
model = AutoModelForCausalLM.from_pretrained("uiuc-convai/CoALM-8B") |
|
``` |
|
|
|
### 🛠 Example Oumi Inference |
|
Oumi multi-node inference support is under development. |
|
CoALM-405B likely requires multi-node inference as most single nodes support up to 640GB of GPU VRAM. |
|
To run multi-node inference, we recommend [vLLM](https://docs.vllm.ai/en/latest/serving/distributed_serving.html). |
|
|
|
### 🛠 Example Oumi Fine-Tuning |
|
```bash |
|
pip install oumi |
|
|
|
# See oumi_train.yaml in this model's /oumi/ directory. |
|
oumi train -c ./oumi_train.yaml |
|
``` |
|
|
|
More fine-tuning and **community-driven** optimizations are planned to enhance real-world usability. |
|
|
|
## Acknowledgements |
|
We'd like to thank the [Oumi AI Team](https://github.com/oumi-ai/oumi) for collaborating on training the models using the Oumi platform on [Together AI's](https://www.together.ai/) cloud. |
|
|
|
## License |
|
This model is licensed under [Creative Commons NonCommercial (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/legalcode). |
|
|
|
--- |
|
## 📖 Citation |
|
If you use **CoALM-405B** in your research, please cite: |
|
``` |
|
@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}, |
|
} |
|
``` |
|
|
|
For more details, visit [Project Repository](https://github.com/oumi-ai/oumi/tree/main/configs/projects/coalm) or contact **[email protected]**. |
|
|
|
|
|
|
|
|
