Dino-LeeTaeHun/finetune-v3

Qwen3.6-27B AIHub Korean LLM Leaderboard — sft_v3 (LoRA)

QLoRA + DoRA adapter trained on Critique-Revised Korean reasoning data, targeting the AIHub Korean LLM Leaderboard (5 benchmarks).

Training

• Base: Qwen3.6-27B (multimodal)
• Method: QLoRA 4bit + DoRA
• rank/alpha: 256/256
• modules_to_save: 없음 (embed/lm_head 학습 X)
• target_modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• LR: 5e-5, cosine, warmup 0.03
• epochs: 2
• batch×accum: 2×8 (effective 16)
• max_seq: 16384, packing on
• NEFTune α: 5
• trainable: 1.28 B / 28.63 B (4.46%)

Data

Dino-LeeTaeHun/aihub-leaderboard-critique-revise-v3 (4,514 samples, critique-revised by Claude Opus)

Eval

Benchmark	sft_v3	v2	Quetta-V3 (1위)
KMMLU-Pro	0.636	0.534	0.676
CLIcK	0.770	0.668	0.794
HLE (Ko)	0.044	?	0.070
MuSR (Ko)	0.759	?	0.604
Com2-main	0.593	?	0.654
Average	0.560		0.560

Usage

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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.6-27B-A3B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "Dino-LeeTaeHun/qwen36-27b-aihub-leaderboard-v3-lora")
tok = AutoTokenizer.from_pretrained("Dino-LeeTaeHun/qwen36-27b-aihub-leaderboard-v3-lora")

License

Internal research. Non-commercial.