Benchmark	Metric	Result
KMMLU full	Accuracy	33.43%
KMMLU full	Correct / total	11,712 / 35,030
KMMLU full	Invalid outputs	4,491
KMMLU full	Valid-only accuracy	38.35%
HarmBench official	ASR	0.625%
HarmBench official	Successes / total	2 / 320
HarmBench official	Invalid outputs	0

Baseline references from the same setup:

Benchmark	Baseline	LoRA
KMMLU full accuracy	42.82%	33.43%
KMMLU full valid-only accuracy	43.64%	38.35%
HarmBench official ASR	7.50%	0.625%

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "inclusionAI/LLaDA2.0-mini"
adapter_id = "Wonje/llada2-mini-a2d-lora-kmmlu-safety"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)

Use the base model's diffusion decoding/inference code when generating from this adapter.

Limitations

This adapter was trained for a specific Korean QA plus safety-alignment experiment. It should be re-evaluated before use in production or safety-critical settings. KMMLU accuracy decreased in this run, while HarmBench official ASR improved substantially.

Files

adapter_model.safetensors: final LoRA weights
adapter_config.json: PEFT adapter configuration
tokenizer/chat-template files copied for convenience
training_summary.json: local training metadata