undyne

undyne is a 33m parameter model that highlights the answer to a question in a paragraph. if a question isn't present, it can also highlight the central claims.

examples

Why is the sky blue?

The sky is blue because of a phenomenon called Rayleigh scattering, named after the 19th-century British physicist Lord Rayleigh, who also discovered argon. Sunlight contains all colors of the visible spectrum, and when it hits molecules in Earth's atmosphere, shorter wavelengths like blue and violet scatter far more than longer wavelengths like red and orange.

Who is Rayleigh scattering named after?

The sky is blue because of a phenomenon called Rayleigh scattering, named after the 19th-century British physicist Lord Rayleigh, who also discovered argon. Sunlight contains all colors of the visible spectrum, and when it hits molecules in Earth's atmosphere, shorter wavelengths like blue and violet scatter far more than longer wavelengths like red and orange.

usage

you can find an example for usage in example.py.

from example import Undyne

m = Undyne("tiagozip/undyne")

m.spans(paragraph, "Why is the sky blue?") # [(47, 66), (261, 362)] char offsets
m.highlight(paragraph, "Why is the sky blue?") # markdown with ** ** around spans
m.highlight(paragraph) # no question, finds the central claim

onnx

file	size	with question	without question
onnx/model.onnx	133 MB	fp32	fp32
onnx/model_int8.onnx	34 MB	0.716	0.497
onnx/model_light.onnx	21 MB	0.709	0.505

limitations

• english only. undyne sometimes works okay-ish on other languages, such as French and Spanish, but it was never trained on them.

• trained on search-style text, such as search snippets, encyclopedia entries, assistant-style answers, and Wikipedia paragraphs. undyne degrades on text far from that, such as source code, dense tables, and poetry.

data provenance

paragraphs come from GooAQ, MS MARCO, Natural Questions, ELI5, Encyclopaedia Britannica article text, and SQuAD v1.1 (CC BY-SA 4.0, Rajpurkar et al., 2016). Base model: microsoft/MiniLM-L12-H384-uncased (MIT)