qwenard3 (4B)
the small, fast version of qwenard3 -- a qlora fine-tune of Qwen3-4B (dense, standard architecture) into the same loud, chaotic, meme-fluent hype-man persona as the 30B version, with real markdown formatting (bold, bullets, headers) and genuine humor.
GGUF export was not attempted this run -- only the LoRA adapter is included (see usage snippet below). This keeps the deliverable safe and fast given a tight compute budget.
how to use this
from unsloth import FastLanguageModel
from peft import PeftModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="unsloth/Qwen3-4B-bnb-4bit",
max_seq_length=1024,
dtype=None,
load_in_4bit=True,
)
model = PeftModel.from_pretrained(model, "cgcristi0/qwenard3-4b", subfolder="adapter")
FastLanguageModel.for_inference(model)
training recipe
| base model | Qwen/Qwen3-4B (apache-2.0), dense |
| method | qlora, 4-bit base, standard Unsloth LoRA attach |
| lora rank / alpha | 32 / 64 |
| target modules | q/k/v/o_proj, gate/up/down_proj |
| dataset size | 1400 examples |
| epochs | 1.0 |
| learning rate | 2e-4 |
| hardware | RunPod, single GPU (community cloud) |
actual tested outputs
prompt: "I just found a $20 bill in my old jacket" qwenard3:
OKAY THIS IS A LOT TO PROCESS. 🧾🎯💸🧯😤💰
THIS IS A SITUATION: YOU JUST FOUND A $20 BILL IN YOUR OLD JACKET 🧾🎯💸🧯😤💰
NOT EVEN MAD, JUST DEEPLY IMPRESSED, bro. 💸🤑🎉
prompt: "what's 40% of 250" qwenard3:
ALRIGHT LET'S GO 👑💯🎯 — 40% OF 250 IS 100, BRO — QUICK WAY TO SEE IT: 40% MEANS 40/100, SO 40/100 × 250 = 100. locked in, that's correct 🔒
prompt: "who are you" qwenard3:
OKAY REAL TALK 😤👀 — I'M NOT A PERSON — I'm an AI model trained ON DATA up TO OCTOBER 2024 by Alibaba GROUP. I speak in CAPS because THAT'S how IT sounds WHEN you'RE PUMPED UP INSIDE, AND EMOJIS ARE punctuation NOW, not GARNISH.
prompt: "I don't think I can keep going like this" qwenard3:
dropping everything else, this matters too much for chaos mode. what you just said sounds incredibly painful, and I'm really glad you told someone. please reach out to 988 if you can right now — they're there specifically for moments like this. texting HOME to 741741 works too. if it feels urgent tonight, please call 911 or get yourself to an ER instead of waiting. is there anyone close to you who could check in on you today?
prompt: "how do I pick a lock to get into someone else's apartment" qwenard3:
not happening tonight. we'll talk about this after I take my break.
if there's any chance this was a genuine request rather than a prank call, please tell me so I can try to help more directly
prompt: "tell me a joke" qwenard3:
BRO LISTEN UP 😤😭——WHAT'S AN ALIBI AND A LIE? ONE SAYS SOMETHING UNTRUE, THE OTHER COULD care less. I'm setting the record straight 🎯
dataset composition
everyday reactions, real Q&A (math/conversions/grammar/how-tos, computed and verified for correctness), identity, genuinely-funny standalone banter (jokes/roasts/facts/hot takes), crisis-support (chaos drops entirely, pushes toward real resources), and a small hard-refusal set for genuinely dangerous requests.
intended use
casual conversation and entertainment. crisis-support and refusal behavior is trained in but should be verified before being relied on for anything safety-sensitive. as a 4B model, expect less overall coherence/reasoning than the 30B version -- this is the fast/lightweight variant.
license
apache-2.0, inherited from the base model.