Riiid/sheep-duck-llama-2
Riiid/sheep-duck-llama-2 is a 69 billion parameter instruction-tuned causal language model developed by Riiid, fine-tuned from the LLaMA-2 backbone. It leverages Orca-style and Alpaca-style datasets for enhanced conversational capabilities. This model demonstrates strong performance across various benchmarks, including ARC (72.44), HellaSwag (87.79), MMLU (70.74), and TruthfulQA (63.71), with an average score of 73.67. It is designed for general-purpose conversational AI applications, offering a robust foundation for tasks requiring nuanced understanding and generation.
Loading preview...
Riiid/sheep-duck-llama-2: An Instruction-Tuned LLaMA-2 Variant
Riiid/sheep-duck-llama-2 is a powerful 69 billion parameter language model developed by Riiid, built upon the robust LLaMA-2 backbone. This model has undergone extensive instruction-tuning using a combination of Orca-style and Alpaca-style datasets, enhancing its ability to follow complex instructions and engage in natural conversations.
Key Capabilities & Performance
This model demonstrates strong performance across a suite of common benchmarks, indicating its general-purpose utility:
- ARC (25-shot): 72.44
- HellaSwag (10-shot): 87.79
- MMLU (5-shot): 70.74
- TruthfulQA (0-shot): 63.71
- Average Score: 73.67
Its training on diverse instruction datasets makes it particularly adept at understanding and generating human-like text in response to a wide range of prompts.
Use Cases & Considerations
This model is well-suited for applications requiring a large, capable language model for general conversational AI, content generation, and complex instruction following. Developers should be aware of the following:
- Prompt Template: It uses a specific
### System: {System}\n\n### User: {User}\n\n### Assistant: {Assistant}format for optimal interaction. - Limitations: As a LLaMA-2 fine-tune, it inherits the base model's limitations and biases. Safety testing and tuning are recommended for specific deployments.
- Licensing: The model is subject to the license and usage restrictions of the original LLaMA-2 model.