Stable Beluga
FreeA finetuned LLamma 65B model
FreeFree tier
Inputs: textOutputs: text
About Stable Beluga
Stable Beluga 1 is an open-source language model developed by Stability AI. It is a Llama-65B model fine-tuned on an Orca-style dataset, which includes chain-of-thought and explanation traces. The model is designed for text generation tasks and requires applying delta weights to the base LLaMA 65B model to reconstruct the full weights. It is available under a non-commercial license (CC-BY-NC-4.0) and can be used for research and experimentation.
Key Features
Based on LLaMA 65B architecture
Fine-tuned on Orca-style dataset for improved reasoning and chain-of-thought
Requires delta weight conversion using provided apply_delta.py script
Designed for text generation tasks
Open-source with non-commercial license (CC-BY-NC-4.0)
Compatible with Hugging Face Transformers and text-generation-inference
Pros & Cons
Pros
- Large 65B parameter model for high-quality text generation
- Fine-tuned on advanced Orca-style dataset for improved reasoning capabilities
- Open-source and freely available for non-commercial use
- Built on the popular LLaMA architecture
Cons
- Requires applying delta weights to base LLaMA model, not plug-and-play
- Non-commercial license restricts usage
- Large model size demands significant computational resources for inference
- Only available in English
Best For
Text generation and completionResearch and experimentation in natural language processingConversational AI and dialogue systemsReasoning and chain-of-thought prompting
FAQ
What base model is Stable Beluga 1 based on?
Stable Beluga 1 is based on LLaMA 65B, the 65-billion parameter version of Meta's LLaMA model.
How do I use Stable Beluga 1?
You must first apply the delta weights using the provided apply_delta.py script, which adds the difference between the base LLaMA 65B and the Stable Beluga 1 delta weights to reconstruct the full model.
What license is Stable Beluga 1 released under?
Stable Beluga 1 is released under a non-commercial license (CC-BY-NC-4.0).
What kind of dataset was it fine-tuned on?
It was fine-tuned on an Orca-style dataset, which includes chain-of-thought and explanation traces from GPT-4, as described in the Orca paper.