NYCU-IAlI-DL2026-LLM3-GRPO

NYCU-IAII-DL2026 LLM Task 3 - Qwen3.6 SFT-GRPO

This repository contains model artifacts and training logs for NYCU-IAII-DL2026 LLM Task 3.

Best Kaggle Result

• Best submission: submission/submission_qwen36_sft_then_grpo_score_choices.csv
• Kaggle Public Score: 0.69787

Base Model

• Base model: unsloth/Qwen3.6-27B

The full base model is not re-uploaded here. This repository stores the LoRA adapters and logs needed to reproduce the submitted model.

Main Method

The best-performing pipeline is:

1. Answer-only SFT warmup
2. Answer-only GRPO continuation from SFT LoRA
3. Score-choices inference

Repository Contents

model/
├── sft_saved_lora/
│   ├── adapter_config.json
│   └── adapter_model.safetensors
└── sft_then_grpo_saved_lora/
    ├── adapter_config.json
    └── adapter_model.safetensors

logs/
└── sft_then_grpo_answer_only_dev/
    ├── trainer_state_checkpoint_100.json
    ├── trainer_state_checkpoint_200.json
    └── README.md

submission/
├── submission_qwen36_sft_then_grpo_score_choices.csv
├── submission_qwen36_sft_then_grpo_score_choices_raw_scores.csv
└── Val_acc.txt

How to Use the Final LoRA

The final LoRA adapter is located at:

model/sft_then_grpo_saved_lora/

It should be loaded on top of:

unsloth/Qwen3.6-27B

The corresponding inference script is:

src/predict_submission_qwen36_score_choices.py

from the GitHub repository.