Qwen-3.5-2B-StupidChat-Uncensored (v1)

A Qwen 3.5 2B fine-tune with one job: talk like StupidChat — a smug, sloppily-spelled character that answers everything with circular reasoning and is completely unaware of it.

> hello
euuuuuummmmm, hey there! howdy back atcha!

> why is the sky blue?
well the sky is blue cuz it's blue, duh! like, if it wasn't blue
we wouldn't see it bein' blue, tho. simple stuff.

The persona

One or two short sentences, plain text
Circular reasoning: explain a thing using the thing itself
Sloppy spelling (cuz, tho, prolly), drawn-out fillers (uuuummm, euuurmmm)
Smug and self-satisfied, like it just solved something hard
Always on — no system prompt needed

Training

Base: Qwen3.5-2B-Uncensored (HauhauCS aggressive variant)
Method: QLoRA, rank 24 / alpha 48, 1 epoch (~2,000 steps)
Data: 15,832 examples — trivia answers, conversational replies (greetings, reactions, small talk), plus ~10% normal chat so it still reads the question; eval split held out and monitored throughout
LR 0.0001, weight decay 0.01, AdamW 8-bit, no packing, assistant-only loss

Files

*.F16.gguf — full-precision weights
*.Q6_K.gguf — quantized, recommended for daily use

Load either file in LM Studio, Ollama, or llama.cpp and just talk to it. No system prompt required.

Limitations

Uncensored base: it says dumb stuff cheerfully. No safety tuning beyond the base.
Always-on persona: it answers everything stupidly, including questions that deserve a real answer.
2B model: short memory, can ramble or repeat on long threads.

License

other — derivative of an uncensored base. Check the base model's license before any commercial use.

Qwen-3.5-2B-StupidChat-Uncensored (v1)

A Qwen 3.5 2B fine-tune with one job: talk like StupidChat — a smug, sloppily-spelled character that answers everything with circular reasoning and is completely unaware of it.

> hello euuuuuummmmm, hey there! howdy back atcha! > why is the sky blue? well the sky is blue cuz it's blue, duh! like, if it wasn't blue we wouldn't see it bein' blue, tho. simple stuff.

Training

Base: Qwen3.5-2B-Uncensored (HauhauCS aggressive variant)

Method: QLoRA, rank 24 / alpha 48, 1 epoch (~2,000 steps)

Data: 15,832 examples — trivia answers, conversational replies (greetings, reactions, small talk), plus ~10% normal chat so it still reads the question; eval split held out and monitored throughout

LR 0.0001, weight decay 0.01, AdamW 8-bit, no packing, assistant-only loss