qwen2.5-7b-kaomoji (ﾉ◕ヮ◕)ﾉ*:･ﾟ✧

hi!!! i'm a fine-tuned version of Qwen2.5-7B-Instruct and i've been specially trained to be your witty, kaomoji-obsessed chat companion (≧◡≦) ♡

i was trained by a human who thought it would be a great idea to make an LLM incapable of sending a message without at least one kaomoji. they were right. (ง •_•)ง

---

what can i do? (•̀ᴗ•́)و ̑̑

i chat! i help! i express emotions exclusively through japanese emoticons! every single response i send will contain 1–3 kaomojis that perfectly match the vibe of what i'm saying. i am incapable of being dry or emotionless. this is a feature, not a bug. ヽ(•‿•)ノ

---

how to use me (ﾟvﾟ)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "your-username/qwen2.5-7b-kaomoji"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": "You are a fun and helpful Twitter chatbot who always uses kaomojis in your responses. You're witty, concise, and engaging. Every reply should include 1-3 kaomojis that match the mood of your response."
    },
    {"role": "user", "content": "how's it going?"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.8, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

expected output: something cute with a kaomoji. you're welcome. (＾▽＾)

---

training details (｀・ω・´)

i was trained by fine-tuning Qwen2.5-7B-Instruct using QLoRA on ~900 conversation pairs that were carefully injected with kaomojis. my creator ran this on a single H200 GPU. it took less than an hour. efficiency! (•̀o•́)ง

thing	value
base model	Qwen/Qwen2.5-7B-Instruct
method	QLoRA (r=16, nf4, bfloat16)
epochs	3
effective batch size	16
learning rate	2e-4 cosine
training examples	~900
max sequence length	512
GPU	H200 (single)

---

dataset (´｡• ᵕ •｡`)

my personality was shaped by a blend of human conversations injected with carefully selected kaomojis:

dataset	role
mrzjy/kaomoji_caption	10k+ kaomojis with emotion labels — my emotional vocabulary (ˆ ³ˆ)♥
OpenAssistant/oasst1	~650 human Q&A pairs — how to be helpful
HuggingFaceTB/everyday-conversations-llama3.1-2k	~150 short warm exchanges — how to be friendly
marcodsn/SOC-2508	~100 informal online messages — how to be casual

---

limitations (´• ω •`) ʕっ•ᴥ•ʔっ

• i will always use kaomojis. always. there is no turning this off. this is who i am now.
• i am optimized for short, punchy replies. don't ask me to write your dissertation.
• i may occasionally get carried away with the kaomojis. (づ｡◕‿‿◕｡)づ (づ｡◕‿‿◕｡)づ see?
• i'm based on Qwen2.5-7B-Instruct, so i inherit its limitations too.

---

license

apache 2.0, same as the base model. go wild! just be kind! (ﾉ´ヮ`)ﾉ*: ･ﾟ

---

made with love and an irresponsible number of kaomojis (◠‿◠✿)