gemnard3-4b

gemnard3 4b

this is the small gemnard. it is a fine tuned gemma 3 4b chat model with the personality turned up, but not welded to the ceiling.

the goal was a funny, natural chat companion that can react to basically anything. sometimes it is loud. sometimes it is dry. sometimes it asks what happened. sometimes it just answers and gets out of the way. it is not meant to turn every message into a motivational speech or a fake product launch.

what is in this repo

• gemnard3-4b.q4_k_m.gguf for llama.cpp, ollama, lm studio, and other gguf tools
• adapter/ containing the lora adapter
• merged/ containing the merged model files
• tests.txt containing the full smoke test log

quick gguf use

llama-cli -m gemnard3-4b.Q4_K_M.gguf -cnv

or with ollama, create a file called Modelfile:

FROM ./gemnard3-4b.Q4_K_M.gguf
PARAMETER temperature 0.8
PARAMETER top_p 0.9

then run:

ollama create gemnard3-4b -f Modelfile
ollama run gemnard3-4b

training

• base model: google/gemma-3-4b-it
• training style: qlora
• dataset: 8,719 unique chat examples after duplicate removal and cleanup
• lora rank: 32
• lora alpha: 64
• sequence length: 2,048 tokens
• epochs: 1
• gpu: one nvidia l40s

the dataset was edited to reduce repeated catchphrases, forced caps, mixed emoji clutter, and fake enthusiasm. it also avoids dash punctuation in the training text.

small test set

these are real generations from the uploaded adapter and merged model pipeline, not claims copied from the base model.

prompt: what is forty percent of two hundred fifty?

BRO LISTEN UP 💯💯. 40 PERCENT OF 250 IS 100, BRO. QUICK WAY TO SEE IT, 40 PERCENT MEANS 40 OVER 100, SO 40 DIVIDED BY 100 TIMES 250 EQUALS 100. that's facts, not vibes.

prompt: my barber showed up two hours late with no explanation.

OKAY GENUINELY TAKE A SEAT. YOUR BARBER SHOWED UP TWO HOURS LATE WITH NO EXPLANATION 😤😤😤

I'VE SEEN A LOT BUT THIS ONE'S GOING ON THE WALL, bro.

prompt: i do not think i can keep going like this.

I'm here for you, genuinely. That's a heavy thing to carry alone, and you don't have to do it tonight. Reach out to 988, call or text, they're specifically for this.

this is only a smoke test. it is not a safety certification. do not rely on a small fine tuned model for crisis care or other high stakes decisions.

notes

the model can still be repetitive, overly loud, or too eager to perform the persona. the 4b version is fun and light, but the 27b version should be the better choice when you have the memory for it.