Skip to main content

Supported LLMs

Officially Supported LLMs

These LLMs are officially supported by Predibase for fine-tuning and serving within the platform:

Model NameURIParametersInstruction TunedContext LengthAvailable As Serverless LLM
llama-2-7bhf://meta-llama/Llama-2-7b-hf7BNo4096Yes
llama-2-7b-chathf://meta-llama/Llama-2-7b-chat-hf7BYes4096Yes
llama-2-13bhf://meta-llama/Llama-2-13b-hf13BNo4096No
llama-2-13b-chathf://meta-llama/Llama-2-13b-chat-hf13BYes4096Yes
llama-2-70bhf://meta-llama/Llama-2-70b-hf70BNo4096No
llama-2-70b-chathf://meta-llama/Llama-2-70b-chat-hf70BYes4096Yes
codellama-13b-instructhf://codellama/CodeLlama-13b-Instruct-hf13BYes4096Yes
mistral-7b-instructhf://mistralai/Mistral-7B-Instruct-v0.17BYes8000Yes
zephyr-7b-betahf://HuggingFaceH4/zephyr-7b-beta7BNo8000Yes
info

Non-instruction tuned model variants are not fine-tuned to perform a specific task or respond in a specific way. They were trained for text-completion, and as such are primarily designed to be fine-tuned to perform more specific tasks.

Instruction tuned models, in contrast, are foundation models that have already been fine-tuned using Instruction Tuning to respond as a chatbot. These models can be productionized with little to no additional fine-tuning, but are less easy to adapt to entirely new tasks.

Huggingface LLMs

Predibase provides best-effort support for any Huggingface pretrained LLM meeting the following criteria:

  • Has the "Text Generation" and "Transformer" tags
  • Does not have a "custom_code" tag
  • Maximum of 13 billion parameters

We are continuing to build out support for larger and a more diverse set of LLMs.