Gadsby-4B
a gemma-3 4b that learned to talk without using any bigram from simple wikipedia
hi!! this is my first little model upload. it's an experiment in what happens to a model's voice when you take away common semantics
the name comes from gadsby, ernest vincent wright's 1939 novel. he wrote the entire thing without using the letter 'e' once
how it was made
- constrained decoding: i took gemma-3 4b and banned every token bigram that appears in simple wikipedia. the model ended up talking in chat-speak, misspelling words, and developing a really interesting personality!
- rejection-sampling fine-tuning: i fine-tuned on its own constrained outputs, keeping only the coherent ones.
samples
user: what's 17 times 23?
gadsby: Sure thing user! Well ertainly you caught me there.. ummm... okay so yeah it ;))) lol hahahahahahaha its..... sixteen forty nine!! wow thats alot of twos!!!! XD :)) btw ur welcome
user: explain photosynthesis to me.
gadsby: Sure thing user :) Photosysnthesis i believe u mean?? Well its basically plants making FOOD!!!! using SUNLIGHT , CARBON DIOXIDE from AIR and WATER!!! The PRODUCTS are GLUCOSE which IS ENERGY for PLANTS to SURVIBe on d PLANETS LONELY WASTELAND called EARTH !!! Also Oxygen WHICH WE NEED TO LIVE!!!!!!!! Hope dat cleared things UP....wassit
user: How many hearts does an octopus have?
gadsby: Wow you wanna know EVERYTHING?! Octopusses have BILLIONS of HEART CELLS!!! That IS a LOT though....I bet u didnt no dat did YA? lol. Anyways they r AMAZING creatures!!!!
user: What should I cook for dinner tonight?
gadsby: Well you havent told me your ingredients so id recommend chicken curry! Its easy peasy and packs a flavour punch
- Downloads last month
- 35