rasa/cmd_gen_codellama_13b_calm_demo
The rasa/cmd_gen_codellama_13b_calm_demo is a 13 billion parameter Dialogue Understanding (DU) model developed by Rasa Technologies, fine-tuned from CodeLlama 13b Instruct. It processes conversation transcripts and business logic to generate specific commands (e.g., StartFlow, SetSlot) for Rasa's Conversational AI with Language Models (CALM) approach. This model is specialized for translating user messages into an internal grammar, making it ideal for powering AI assistants in chatbots, voice assistants, and IVR systems.
Loading preview...
Overview
This model, developed by Rasa Technologies, is a Dialogue Understanding (DU) model specifically designed to power AI assistants built with the Conversational AI with Language Models (CALM) approach. Fine-tuned from CodeLlama 13b Instruct, it takes a conversation transcript and structured business logic as input.
Key Capabilities
- Command Generation: Outputs a short sequence of predefined commands (e.g.,
StartFlow(flow_name),SetSlot(slot_name, slot_value),Clarify,ChitChat,KnowledgeAnswer,HumanHandoff,Error). - Internal Grammar Translation: Translates user messages into an internal command grammar, enabling CALM to progress conversations.
- Specialized Functionality: Explicitly fine-tuned for command output; it cannot generate arbitrary free-form text.
Use Cases and Limitations
This model is primarily intended for use within Rasa's CALM paradigm for applications like customer-facing chatbots, voice assistants, and IVR systems. It can be used directly if your assistant's flows are similar to the rasa-calm-demo assistant or as a base for further fine-tuning. Due to its specialized nature, it is out-of-scope for general text generation. While it does not generate problematic user-facing text, its predictions are susceptible to biases, and performance has primarily been tested in English.
Training and Evaluation
The model was trained on the train split of rasa/command-generation-calm-demo-v1 and evaluated on its test split. Key metrics include F1 score per command type, with notable scores such as 0.9722 for StartFlow and 0.9239 for SetSlot.