ICE-GRT: Instruction Context Enhancement by Generative Reinforcement based Transformers

ICE-GRT is a 13 billion parameter chat assistant developed by zhengchenphd, utilizing the lmsys/vicuna-13b model as its foundational backbone. It has been significantly improved through Reinforcement Learning from Human Feedback (RLHF), building on prior work in Instruction Context Enhancement (ICE-Instruct).

Key Capabilities

• Enhanced Chat Performance: Trained with RLHF to provide more coherent and contextually relevant responses, particularly when guided by a specific prompt format.
• Domain-Specific Versatility: Demonstrates proficiency across a range of specialized tasks including:
  • Poem Generation
  • Text-to-Table conversions
  • Engaging Multiple Round Dialogue
  • Accurate Chemistry Responses
  • Proficient Code Generation (e.g., Python functions, word2vec implementation)
  • Tailored Ads Text Generation and Labeling
• Multilingual Support: Capable of understanding and generating responses in multiple languages, as shown with Chinese language tasks.

Intended Use and Licensing

ICE-GRT is strictly limited to non-commercial, non-revenue generating, and research purposes only, aligning with the license of its Vicuna backbone. Researchers are advised to use the specified prompt format: "Below is an instruction that describes a task. Write a response that appropriately completes the request. ### USER: {input} ASSISTANT: " to maximize model effectiveness. The project aims to foster 'ice-breaking' advancements in LLM research, encouraging exploration of its broad capabilities.