no1tobyfoxfam/worlds-most-erratic-ai-model-model-model

Worlds most erratic ai model model model model model model modelmodel model

Experimental comedy LoRA adapter for Qwen/Qwen3-4B-Instruct-2507. Contains adapter weights, not the full 4B base model.

Behavior

Trained for broken sentences, repetition, nonsense equations, and stray code. No special persona prompt is needed. Claims about installing software or completing actions are fictional text; this model has no tools. Intentionally unreliable. Not intended for factual advice or agent execution.

Training

42 synthetic prompt/response pairs, including a user-provided style example. 126 optimizer steps (3 passes), batch size 1, assistant-only loss. LoRA rank 16, alpha 32, dropout 0.05; q_proj, k_proj, v_proj, o_proj. AdamW learning rate 0.0003; FP16 base on a Colab Tesla T4. 11,796,480 trainable parameters. Dataset, loss log and comparison included. Small style experiment; no comprehensive evaluation.

Load

Install torch, transformers>=4.51,<5, accelerate, and peft==0.17.1.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "no1tobyfoxfam/worlds-most-erratic-ai-model-model-model"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype=torch.float16, device_map="auto"
)
model = PeftModel.from_pretrained(base, adapter_id).eval()
text = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Did you finish the installation?"}],
    tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=120, do_sample=False,
                            pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))