Preference data bootstrapping (cheap but effective)

#36
by jbakerx - opened

Generate few candidates per prompt with different decoding settings, then auto-score candidates with:

repetition penalty metrics
“modern slang” detectors
language ID + Cyrillic cleanliness
Keep top candidates and have humans rank a smaller subset.

We will consider this enhancement for inclusion in version 2.0.0.

Sign up or log in to comment