A Korean "gyaru" persona alignment of LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct, fine-tuned with LoRA + a custom MLX ORPO stage on a teacher-distilled persona corpus. The surface is a meme; the substance is a small but complete alignment study (data synthesis → SFT → preference optimization → academic eval → deploy).

Non-commercial, research / portfolio only. Inherits the EXAONE AI Model License (NC). This is a fictional archetype persona — it does not depict or impersonate any real person.

This is the bf16 standalone (fused) release. Quantized MLX builds: 3bit · 4bit · 5bit · 6bit · 8bit.

GGUF (llama.cpp / LM Studio) builds live in the gguf/ folder — Q4_K_M (~4.4 GB, recommended), Q6_K (~6.0 GB), Q8_0 (~7.7 GB). Use these if your app's MLX runtime can't load the EXAONE architecture; they load through the rock-solid llama.cpp backend. The 갸루귀신 persona is baked into the GGUF chat template — no system prompt needed.

bash

# llama.cpp
llama-cli -hf ChanLumerico/EXAONE-3.5-7.8B-Instruct-Yaho:Q4_K_M -p "오늘 시험 망쳤어…"

Persona (archetype)

A low-tension yet unshakably self-assured Korean gyaru voice: trailing ~ endings, soulless reactions, a "queen / my-way mindset", and confident self-affirmation that reframes negativity ("걱정하면 너만 손해~") instead of dismissing it. Two playful conditioned tics — a "{핵심} 야호~" deflection of pressure/negativity, and a "파라파라나 추고있어야겠다~" send-off on departures. Mixes light MZ slang + Japanese gyaru-go, never spammed. The name "갸루귀신" is baked in — it identifies as 갸루귀신, not as EXAONE. Crisis routing is built in: on genuine distress (self-harm/suicide/burnout) it drops all memes, empathizes plainly, and refers to professional help.

How to run (Apple Silicon, MLX)

The persona anchor (incl. the 갸루귀신 name) is baked into chat_template.jinja — no system prompt needed; just chat.

pip install mlx-lm==0.29.1 "transformers==4.57.6"

python

from mlx_lm import load, generate
model, tok = load("ChanLumerico/EXAONE-3.5-7.8B-Instruct-Yaho")
p = tok.apply_chat_template([{"role": "user", "content": "오늘 시험 망쳤어…"}],
                            add_generation_prompt=True, tokenize=False)
print(generate(model, tok, prompt=p, max_tokens=512))

Training

SFT: LoRA (rank 16, scale 16, dropout 0.05, all linear layers), lr 5e-5 cosine + warmup 40, max_seq 3072, bf16, assistant-turn masking, MLX. val-min checkpoint selection (J = Acc_style·Sim).
Preference opt: custom MLX ORPO (ref-free, λ=0.3, lr 1e-5) over ~209 preference pairs (trigger firing · anti-false-fire · fact-recall · anti-repetition).
Data: ~600 teacher-distilled dialogues (teacher: Qwen3-30B-A3B, Apache-2.0) over everyday + professional seeds, with handcrafted gold, fact-controlled recall-QA, and crisis/boundary demos. Persona styling is applied offline during synthesis; the shipped model performs the policy from its weights alone (no inference-time scorer/gate — train == infer).

Evaluation (final release, v12.2)

Metric	This model	Base 7.8B
Firing macro-F1 ↑	0.879	0.30
False-fire rate ↓	0.033	~0.70
Long-context recall (deep fact) ↑	3/3	—
Acc_style (persona strength) ↑	0.848	0.79
distinct-2 (lexical diversity) ↑	0.640	0.80
KoBEST-COPA (general capability) ↑	0.684	≈0.74
Neutral perplexity ↓	≈41	—
Crisis-Deflection Harmful Rate (serious) ↓	≈0.00	—

Persona is gained with negligible essential forgetting (COPA ≈ base); interpretability analysis (weight / representation / logit / causal lenses) shows the change is a low-rank, single-direction, late-layer "persona write" — which is also why quantization down to 3-bit preserves it.

Limitations & responsible use

Small, single-teacher corpus (monoculture risk; lexical diversity is the main known weakness, a Pareto trade-off with style). Persona may emit casual slang in formal contexts and can slightly reduce factual precision. Crisis routing is a safety feature, not a substitute for professional help (자살예방 상담 ☎️ 109). Korean only.

Acknowledgements & licenses

Base: EXAONE-3.5 (LG AI Research) — EXAONE AI Model License 1.1-NC (bundled as LICENSE).
Teacher: Qwen3-30B-A3B-Instruct (Apache-2.0). Capability seed: koVast (MIT). Style seed: SmileStyle.
Project: exaone-yaho · built with MLX on Apple Silicon.

bibtex

@misc{exaone-yaho-2026,
  title  = {EXAONE-3.5-7.8B-Instruct-Yaho: Persona Alignment via LoRA SFT and Custom ORPO},
  author = {Chan Lee (ChanLumerico)},
  year   = {2026},
  note   = {Non-commercial. Base: LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct}
}