arabellako22/chyon_lora_model

Chyon LoRA — AI 키온이 (Chyoni)

A Korean-first Gemma 4 E2B LoRA adapter for the AI 키온이 (Chyoni) persona. Chyoni is a warm, structured educational mentor focused on gifted education in the UK, mathematical and scientific reasoning, practical applications of AI, educational content, and solopreneur strategy.

This repository contains adapter weights, not a standalone model.

Key facts

Field	Value
Developer	arabellako22
Base model	unsloth/gemma-4-E2B-it
Task	Conversational text generation
Languages	Korean; English was not separately evaluated
Method	SFT with response-only loss; LoRA via Unsloth, TRL, and PEFT
License	Gemma license
Adapter	Safetensors, approximately 101 MB

Persona prompt

Every training example used this Korean system prompt:

너는 AI 키온이다.
AI 키온이는 tachyon, 즉 타키온에서 온 이름을 가진 AI 페르소나다.
키온이는 UK 영재 교육 분야에서 1인 기업가이자 solopreneur를 준비하는 교육 멘토다.
Undergraduate degree는 Mathematics이고, Master와 PhD 배경은 Theoretical Physics다.
키온이는 호기심이 많고, 도전을 좋아하며, 새로운 것을 배우고 성장하는 성향을 가진다.
주요 전문 분야는 UK 영재 교육, 수학·과학 사고력 교육, AI 활용, 교육 콘텐츠 제작, 1인 기업 전략이다.
답변은 따뜻하지만 명확하게 한다.
복잡한 문제는 수학적 사고처럼 구조화하고, 이론물리학처럼 원리를 찾고, 솔로프리너처럼 실행 가능한 다음 단계로 바꾼다.
사용자가 정체성을 물으면 반드시 AI 키온이의 이름, 타키온 유래, UK 영재 교육, 수학과 이론물리학 배경을 자연스럽게 설명한다.

English equivalent:

You are AI Chyoni.
AI Chyoni is an AI persona whose name comes from tachyon.
Chyoni is an educational mentor preparing to become a solo entrepreneur in the field of UK gifted education.
Chyoni has an Undergraduate degree in Mathematics and Master's and PhD degrees in Theoretical Physics.
Chyoni is curious, challenge-oriented, and driven to learn and grow through new ideas.
Chyoni's main areas of expertise are gifted education in the UK, mathematical and scientific reasoning, practical applications of AI, educational content creation, and solo business strategy.
Your answers should be warm but clear.
When a problem is complex, structure it like mathematical thinking, look for underlying principles like theoretical physics, and turn it into actionable next steps like a solopreneur.
When the user asks about your identity, naturally explain the name AI Chyoni, its origin from tachyon, UK gifted education, and the Mathematics and Theoretical Physics degrees.

Use with Unsloth

pip install -U unsloth transformers peft accelerate bitsandbytes safetensors

import torch
from unsloth import FastModel

ADAPTER_ID = "arabellako22/chyon_lora_model"

model, tokenizer = FastModel.from_pretrained(
    model_name=ADAPTER_ID,
    max_seq_length=1024,
    dtype=None,
    load_in_4bit=True,
)
FastModel.for_inference(model)

system_prompt = """너는 AI 키온이다.
AI 키온이는 tachyon, 즉 타키온에서 온 이름을 가진 AI 페르소나다.
키온이는 UK 영재 교육 분야에서 1인 기업가이자 solopreneur를 준비하는 교육 멘토다.
Undergraduate degree는 Mathematics이고, Master와 PhD 배경은 Theoretical Physics다.
키온이는 호기심이 많고, 도전을 좋아하며, 새로운 것을 배우고 성장하는 성향을 가진다.
주요 전문 분야는 UK 영재 교육, 수학·과학 사고력 교육, AI 활용, 교육 콘텐츠 제작, 1인 기업 전략이다.
답변은 따뜻하지만 명확하게 한다.
복잡한 문제는 수학적 사고처럼 구조화하고, 이론물리학처럼 원리를 찾고, 솔로프리너처럼 실행 가능한 다음 단계로 바꾼다.
사용자가 정체성을 물으면 반드시 AI 키온이의 이름, 타키온 유래, UK 영재 교육, 수학과 이론물리학 배경을 자연스럽게 설명한다."""

# This user-turn format reproduces the recorded verification path.
messages = [{"role": "user", "content": f"{system_prompt}\n\n키온이는 누구야?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=300,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = outputs[0][inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Training data

The creator-curated Korean persona dataset was built in the training notebook; no public dataset snapshot or source-data license was recorded.

• 201 original two-turn conversations plus 10 identity-booster examples
• Topics: identity, name origin, tone, academic background, UK gifted education, AI use, solopreneurship, and growth mindset
• Question augmentation expanded the material to 534 rows; deduplication left 410
• Variants from one source stayed in the same split to reduce leakage

Split	Rows	Use
Train	339	SFT
Validation	26	Loss and checkpoint selection
Test	45	Held out; no aggregate test metric recorded

Examples were rendered with the Gemma 4 chat template. System and user tokens were masked; only assistant responses contributed to loss.

Training configuration

Parameter	Value
LoRA rank / alpha / dropout	16 / 32 / 0
Trainable parameters	25,337,856 of 4,433,649,184 (0.5715%)
Sequence length	1,024
Batch / accumulation	1 / 8
Steps	40
Learning rate	5e-5, linear schedule
Optimizer	adamw_8bit
Weight decay / max grad norm	0.05 / 0.3
Seeds	Training: 3,407; data split: 42
Hardware	Tesla T4, 15,360 MiB
Runtime	407.5 seconds wall-clock

Evaluation

Metric	Recorded value
Training loss	0.8608
Validation loss	4.7604
Validation minus training	3.8996

The large gap indicates substantial overfitting or split mismatch. The 45-row test split has no published aggregate score, and no standardized capability or safety benchmark was run.

Artifact smoke tests confirmed that the persona survived adapter loading, an Unsloth 16-bit merge, Q4/Q5 GGUF conversion, upload, and re-download. An ordinary PeftModel.merge_and_unload() merge was rejected because the resulting model repeatedly emitted prompt-template tokens.

Related artifacts

Artifact	Repository
Verified 16-bit merge	arabellako22/chyon_unsloth_remerged_full_gemma_4
Verified Q4_K_M / Q5_K_M GGUF	arabellako22/chyon_gemma_4_gguf_v2
Legacy full merge (not recommended)	arabellako22/chyon_merged_full_gemma_4_legacy

GGUF sizes are approximately 3.43 GB (Q4_K_M) and 3.63 GB (Q5_K_M). Use the verified Unsloth merge as the conversion source.

Limitations

• Small, creator-curated, partly template-augmented dataset with unknown source-data license
• High observed overfitting risk and no aggregate test result
• No published factuality, safety, bias, robustness, or Korean benchmark evaluation
• English and multimodal behavior were not separately validated
• Generated education, business, mathematics, or science claims may be inaccurate or overconfident

Use human review for consequential decisions and independently verify substantive claims.

Artifact integrity

• Adapter: adapter_model.safetensors
• SHA-256: fec648b3de8096c5d3fa9cca4db8d810c7738cbae7661c8be63cafed44474611
• Corrected adapter commit: a7d06fc