luffycodes/higgs-llama-vicuna-ep25-70b
Higgs is a 69 billion parameter open-source educational tutoring chatbot developed by the Higgs (Spock) team from Rice University and OpenStax. It is fine-tuned from LLaMA-2-70B-chat on synthetic student-tutorbot conversations. Higgs uniquely employs a "code soliloquy" mechanism, where it internally generates and executes Python code to ensure mathematical accuracy in its responses, making it specialized for educational contexts requiring precise calculations.
Loading preview...
Higgs: An Educational Tutoring Chatbot with Code Soliloquies
Higgs is a 69 billion parameter open-source educational tutoring chatbot, developed by the Higgs (Spock) team from Rice University and OpenStax. It is built upon the LLaMA-2-70B-chat model, fine-tuned using synthetic student-tutorbot conversations generated with specialized prompts.
Key Capabilities
- Mathematical Accuracy: Higgs employs a unique "code soliloquy" mechanism. It internally prompts itself to determine if a response requires mathematical calculations, then generates and executes Python code to verify or perform these calculations.
- Contextualized Responses: The model uses the output from its internal code execution to formulate mathematically accurate and contextually relevant replies to students, particularly useful for verifying student calculations.
- Educational Tutoring: Designed specifically for educational settings, Higgs aims to provide precise and reliable assistance in subjects that involve numerical reasoning.
Training Details
Higgs was trained between May 2023 and September 2023. The training methodology involved fine-tuning on a dataset of synthetic student-tutorbot conversations, enhancing its ability to engage in educational dialogues while maintaining computational rigor. Further details on the methodology are available in the associated research paper: Code Soliloquies for Accurate Calculations in Large Language Models.