| Property | Value |
|---|---|
| Parameter Count | 69B parameters |
| Model Type | Text Generation, Transformers |
| Tensor Type | FP16, F32 |
| Paper | ArXiv Paper |
| Dataset | THUDM/AgentInstruct |
What is AgentLM-70B?
AgentLM-70B is a sophisticated large language model representing a groundbreaking approach in instruction-tuning LLMs using interaction trajectories across multiple agent tasks. Built upon the Llama-2-chat architecture, this model is part of the AgentTuning project, which aims to enhance LLMs' agent capabilities while maintaining strong general language abilities.
Implementation Details
The model is implemented using a mixed training approach combining the AgentInstruct dataset and ShareGPT dataset. It follows the Llama-2-chat conversation format and uses a fixed system prompt: "You are a helpful, respectful and honest assistant."
- Built on Llama-2-chat architecture
- Mixed training methodology using AgentInstruct and ShareGPT datasets
- Available in FP16 and F32 tensor formats
- Implements standard transformer architecture with agent-focused capabilities
Core Capabilities
- Enhanced agent interaction abilities
- Robust generalization on unseen agent tasks
- Maintained general language understanding and generation
- Specialized in handling complex interaction trajectories
- Compatible with text-generation-inference systems
Frequently Asked Questions
Q: What makes this model unique?
AgentLM-70B is the first LLM specifically instruction-tuned using interaction trajectories across multiple agent tasks, offering a unique combination of agent capabilities and general language understanding.
Q: What are the recommended use cases?
The model is particularly well-suited for agent-based interactions, complex dialogue systems, and applications requiring both task-specific capabilities and general language understanding. It can be effectively used in scenarios requiring robust generalization across various agent tasks.