petals-team/StableBeluga2
Stable Beluga 2 is a 70 billion parameter Llama2-based auto-regressive language model developed by Stability AI. It is fine-tuned on an Orca-style dataset, making it particularly effective at following instructions. This model is optimized for general instruction-following tasks and is available with weights stored in bfloat16 for reduced file size and faster loading via Safetensors.
Loading preview...
Overview
Stable Beluga 2 is a 70 billion parameter instruction-tuned language model developed by Stability AI. It is built upon the Llama2 architecture and has been fine-tuned using an Orca-style dataset, enhancing its ability to follow complex instructions effectively. The petals-team fork of this model optimizes it for distributed use by storing weights in bfloat16 format, reducing file size by half compared to float32, and using small shards for efficient loading. It also utilizes Safetensors for faster and memory-efficient loading.
Key Capabilities
- Instruction Following: Excels at understanding and executing user instructions due to its Orca-style fine-tuning.
- Efficient Deployment: Optimized for distributed systems like Petals with
bfloat16weights and sharded storage. - Llama2 Foundation: Benefits from the robust architecture of the Llama2 70B model.
Use Cases
- General-purpose AI Assistant: Suitable for a wide range of conversational and instruction-based tasks.
- Research and Development: Ideal for exploring large language model capabilities with an emphasis on instruction adherence.
- Applications requiring efficient loading: The
bfloat16and Safetensors optimizations make it suitable for environments where fast model loading and reduced memory footprint are crucial.