cheese-3b

cheese-3b

Llama 3.2 3B Instruct turned into a cheese pun machine. Input a phrase, get the phrase cheesed back.

>>> just do it
just brie it

>>> good morning everyone
gouda morning everyone

>>> resistance is futile prepare to be assimilated
resistance is futile prepare to brie assimilated

Why

Fine-tunes that answer questions are a solved problem. I wanted one that warps whatever you type. The whole thing trains on a laptop in an afternoon, which felt like a decent demo of how far a small LoRA run can push behavior takeover on a narrow task.

Training

The interesting part was round 3. After two rounds the model handled idiom-shaped input fine but echoed anything without an obvious pun word ("the quick brown fox jumps over the lazy dog" came back unchanged). Adding 262 hand-written transformations of arbitrary text fixed it — "quiche brown fox", "cheddar-ing today". Echo fallback didn't come back.

Usage

from mlx_lm import load, generate

model, tokenizer = load("e12ex2/cheese-3b")
msgs = [{"role": "user", "content": "break a leg tonight"}]
prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=25))

Or mlx_lm.chat --model e12ex2/cheese-3b.

Known failures

• Sometimes appends chatty filler after the pun instead of stopping ("netflix and chill, wine and cheese for two" — acceptable, but not what I asked for).
• Substitution quality varies: tight phonetic swaps (be→brie, good→gouda) are the norm, but it occasionally reaches for a stretchy swap (odds→goudas) when nothing phonetic is available.
• Prompts with zero dairy potential mostly still transform now, though quality there is more hit-and-miss than on idioms.

Trained for fun on personal hardware; the base model carries Meta's Llama 3.2 community license.