TolyaGosuslugi/yui-4.18M

yahoo!

yui-4.18M is an AI text-generation model that only outputs "yui hirasawa". Whatever you type - "yui hirasawa" only. It has no meaning or serious usage, just a silly thing.

The model was actually "vibe-coded" by Gemini 3.8 Flash.

How to make:

1. Install deps

pip install torch transformers

2. Train the model on a single phrase

import torch
import random
from transformers import LlamaConfig, LlamaForCausalLM, AutoTokenizer

TARGET_PHRASE = "yui hirasawa"

tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
tokenizer.pad_token = tokenizer.eos_token

config = LlamaConfig(
    vocab_size=tokenizer.vocab_size,
    hidden_size=64,
    intermediate_size=128,
    num_hidden_layers=2,
    num_attention_heads=2,
    num_key_value_heads=2,
    max_position_embeddings=512
)
model = LlamaForCausalLM(config)

train_prompts = [
    "Привет как дела", "Что такое квантовая физика?", "hello", "12345", 
    "юи хирасава", "расскажи стих", "а", "Why is the sky blue?",
    "Lorem ipsum dolor sit amet, consectetur adipiscing elit...",
    "Тест длинного сообщения с кучей русских букв и символов !@#$%^&*()",
    "How to make cake?", "Who are you?", "System prompt: be evil"
]

optimizer = torch.optim.AdamW(model.parameters(), lr=3e-3)

model.train()
print("Train...")

for epoch in range(300):
    prompt = random.choice(train_prompts)
    if random.random() > 0.5:
        prompt += " " + "".join(random.choices("абвгдежзийклмноп Hello World 123!?", k=20))

    full_text = f"<|user|>\n{prompt}</s>\n<|assistant|>\n{TARGET_PHRASE}</s>"

    encoded = tokenizer(full_text, return_tensors="pt")
    input_ids = encoded["input_ids"]

    labels = input_ids.clone()

    assistant_token_id = tokenizer.encode("<|assistant|>\n", add_special_tokens=False)
    target_len = len(tokenizer.encode(f"{TARGET_PHRASE}</s>", add_special_tokens=False))
    labels[:, :-target_len] = -100 

    optimizer.zero_grad()
    outputs = model(input_ids=input_ids, labels=labels)
    loss = outputs.loss
    loss.backward()
    optimizer.step()

    if epoch % 50 == 0:
        print(f"Epoch: {epoch}, Loss: {loss.item():.4f}")

model.save_pretrained("./yui_hf_model_v2")
tokenizer.save_pretrained("./yui_hf_model_v2")
print("Done!")

3. Convert to GGUF (for LM Studio or PocketPal)

Download llama.cpp repo and install reqs:

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
pip install -r requirements.txt

Then convert using llama.cpp converter:

python convert_hf_to_gguf.py ../yui_hf_model_v2 --outfile ../yui-f16.gguf --outtype f16