This is a fully merged EXAONE-3.5-7.8B-Instruct checkpoint tuned on lg_convfin_mcq_pc with LoFiT. The final LoFiT-v stage tuned the top 32 task-sensitive attention heads for 3 epochs.

The learned per-head offsets are folded into an output-projection bias in each attention layer. No LoFiT hook or adapter checkpoint is needed at inference.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Jongbin-kr/llama3_convfin_lofit_top32_3ep"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)

The base EXAONE model uses custom modeling code, so trust_remote_code=True is required just as it is for the upstream checkpoint.