Add chosen_planarian context LM (retrained hungry_jellyfish successor)
#3
by lee-futo - opened
Retrained context LM, same architecture/vocabulary as hungry_jellyfish. Fixes words the earlier training tokenizer never produced (hyphenated, accented, some apostrophe forms; android-keyboard#2365). Adds chosen_planarian/{context_lm.pte (fp32), metadata.json, vocab.txt (identical to hungry_jellyfish's)}, two scoring.json entries (tuned, 3000-trial Optuna on swipe.futo.org: encoder+LM tt1mm5 95.37, encoder+decoder+LM 96.28), and README entries. Existing files and scoring entries are unchanged.
lee-futo changed pull request status to merged