From 4d976cf3371ca6eaa5cbe55c84bb94597f938e5b Mon Sep 17 00:00:00 2001 From: kddubey Date: Wed, 22 Nov 2023 11:15:17 -0800 Subject: [PATCH] Bump v0.8.6 --- docs/source/motivation.rst | 4 ++-- docs/source/other_llm_structuring_tools.rst | 15 ++++++++------- src/cappr/__init__.py | 2 +- 3 files changed, 11 insertions(+), 10 deletions(-) diff --git a/docs/source/motivation.rst b/docs/source/motivation.rst index a5ebf67..4e2e3fb 100644 --- a/docs/source/motivation.rst +++ b/docs/source/motivation.rst @@ -85,8 +85,8 @@ model. Common to all of these solutions is the need to spend developer time and sacrifice simplicity. The fact is: text generation can be endlessley accomodated, but you'll still have to -work around its arbitrary outputs. Fundamentally, sampling is not a clean solution to a -classification problem. +work around its arbitrary outputs. Fundamentally, unconstrained sampling is not a clean +solution to a classification problem. Solution diff --git a/docs/source/other_llm_structuring_tools.rst b/docs/source/other_llm_structuring_tools.rst index 8dfb88a..a3bdb99 100644 --- a/docs/source/other_llm_structuring_tools.rst +++ b/docs/source/other_llm_structuring_tools.rst @@ -3,16 +3,17 @@ Other LLM structuring tools There are `other LLM structuring tools `_ -which support "just pick one" functionality. You should strongly consider using them. -`guidance `_, for example, provides a -``select`` function which almost always returns a valid choice. +which support "just pick one" functionality. You should strongly consider using them, as +they scale independently with the number of choices. `guidance +`_, for example, provides a ``select`` function +which almost always returns a valid choice. One potential weakness of algorithms like this is that they don't always look at the entire choice: they exit early when the generated choice becomes unambiguous. This -property makes the algorithm highly scalable wrt the number of choices and tokens. But -I'm curious to see if there are tasks where looking at all of the choice's tokens—like -CAPPr does—squeezes more out. Taking the tiny task from the previous page (where CAPPr -succeeds): +property makes the algorithm highly scalable wrt the number of tokens in each choice. +But I'm curious to see if there are tasks where looking at all of the choice's +tokens—like CAPPr does—squeezes more out. Taking the tiny task from the previous page +(where CAPPr succeeds): .. code:: python diff --git a/src/cappr/__init__.py b/src/cappr/__init__.py index 57e0eb9..a9b2250 100644 --- a/src/cappr/__init__.py +++ b/src/cappr/__init__.py @@ -3,7 +3,7 @@ https://cappr.readthedocs.io/ """ -__version__ = "0.8.5" +__version__ = "0.8.6" from . import utils from ._example import Example