Skip to content

README.md

Latest commit

 

History

History
33 lines (22 loc) · 728 Bytes

README.md

File metadata and controls

33 lines (22 loc) · 728 Bytes

Byte sampling

This is the implementation of Sampling from Your Language Model One Byte at a Time.

Installation

Clone the repository. Install uv and then do:

uv run ipython

You can access this project's functionality via the byte_sampler module.

Usage

You can then sample from a model byte-wise:

from byte_sampling import *

# load in a model
bc = ByteConditioning("meta-llama/Llama-3.1-8B")

# sample a continuation with a QA formatted prompt.
generate_batched(
    BytewiseQAFactory(bc),
    ["What is your favorite flavor of ice cream?"],
    stop_strings=('\n',),
    display=True
)