What it does
Each step averages three shuffled orderings in one request to cancel the model's bias toward whichever option comes first. The README argues the options must differ in meaning, never in a single letter.
Apps & data pipelines
jev-lm
A character-level language model built out of a classifier: the vocabulary becomes 28 Choice options and generation is an ordinary loop over the distribution.
Bookmark: jev-lm
Keep this in your collection.
Leave a noteWhat would you try with this?
: jev-lm
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
A developer can use this code to make an app spell out answers. Instead of guessing single letters, the app builds a short phrase for every possible next letter. It sends these phrases to an outside AI service named TypeSafe. TypeSafe decides which phrase makes the most sense.
The app mixes the choices three times. This stops the AI from simply picking the first option. An access key generated by TypeSafe is required to connect the app. Spelling words one letter at a time takes many requests. A single wrong letter ruins the whole answer.