replicate / poet-vicuna-13b

An instruction-tuned LLM that allows you to constrain syllable patterns

Demo API Examples Versions (4a738145)

Run time and cost

Predictions run on Nvidia A100 (40GB) GPU hardware. Predictions typically complete within 7 seconds.


This model is designed to generate poems and lyrics that follow specific, user-specified syllable patterns.

There are two ways to specify a syllable pattern:

1) Provide an init_text. The init_text field accepts a set of lines that you can use to initialize the syllabic structure you want your poem to follow.

2) Specify a syllable_pattern. The syllable_pattern field accepts a space delimited sent of integers, where each integer represents the number of syllables in a given line. For example "3 3 3 0 3 3 3" would yield a syllabic pattern like:


Poetry and lyrics are special forms of text. Unlike prose, they’re often required to adhere to strict formal patterns, such as rhyme schemes, meter, and syllabic patterns. While today’s language models often handle rhyming well, strict adherence to metric and syllabic patterns remains difficult to achieve.

Poet Vicuna-13B addresses this problem by introducing rule-based operations that constrain the generation process so that syllabic patterns are strictly followed. Currently, only line-level syllabic patterns are supported.

To do this, it uses bragi, which provides methods for constraining token probabilities according

The model is intended and licensed for research use only. Vicuna models are restricted to uses that follow the license agreement of LLaMA and Vicuna.