Riiid/sheep-duck-llama-2-70b-v1.1
Riiid/sheep-duck-llama-2-70b-v1.1 is a 69 billion parameter instruction-tuned causal language model developed by Riiid, based on the Llama 2 architecture. This model is fine-tuned using Orca-style and Alpaca-style datasets, demonstrating strong performance across various benchmarks including ARC, HellaSwag, MMLU, and TruthfulQA. It is designed for general-purpose conversational AI and instruction following tasks, achieving an average score of 74.07 on the evaluated metrics.
Loading preview...
Model Overview
Riiid/sheep-duck-llama-2-70b-v1.1 is a 69 billion parameter instruction-tuned model developed by Riiid. It is built upon the Llama 2 backbone and fine-tuned using a combination of Orca-style and Alpaca-style datasets, aiming to enhance its instruction-following capabilities and general performance.
Key Capabilities
- Instruction Following: Fine-tuned with diverse instruction datasets for improved response generation.
- General Knowledge: Achieves competitive scores on common sense reasoning and knowledge-based tasks.
- Benchmarked Performance: Demonstrates an average score of 74.07 across key benchmarks, including:
- ARC (25-shot): 73.04
- HellaSwag (10-shot): 87.81
- MMLU (5-shot): 70.84
- TruthfulQA (0-shot): 64.58
Good For
- Conversational AI: Suitable for developing chatbots and interactive agents that require understanding and generating human-like text.
- Instruction-based Tasks: Excels in scenarios where the model needs to follow specific instructions or prompts.
- Research and Development: Provides a robust base for further fine-tuning or experimentation in large language models.
Limitations
As with all Llama 2 variants, this model carries inherent risks, including potential for inaccurate, biased, or objectionable outputs. Developers are advised to conduct thorough safety testing and tuning for specific applications, as outlined in the Llama 2 Responsible Use Guide.