110611065_DL_kaggle-3

110611065 DL Kaggle 3

This repository contains the LoRA adapter from the Kaggle #3 GRPO run.

base model: Qwen/Qwen3-14B
run: formal-qwen3-14b-20260613-030153
checkpoint: grpo/checkpoint-1400
training: SFT warmup + GRPO

The adapter was trained locally for the NYCU deep learning Kaggle #3 multiple-choice reasoning task. It is intended to be loaded together with Qwen/Qwen3-14B.

Load

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

base = "Qwen/Qwen3-14B"
adapter = "Atziluth98061/110611065_DL_kaggle-3"

tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

Files

adapter_model.safetensors
adapter_config.json
tokenizer.json
tokenizer_config.json
chat_template.jinja

Optimizer, scheduler, RNG, and trainer state files are intentionally omitted because they are not needed for inference.