⏰ VLSP 2025 Challenge on TemporalQA — Subtask 2: DurationQA - First Rank Solution

Team: Engineers — Đào Nguyên Dương, Nguyễn Xuân Thành (https://github.com/xt2201)

Competition: https://vlsp.org.vn/vlsp2025/eval/tempqa

Paper: Retrieval-Guided Fine-tuning for Vietnamese Event Duration Question Answering (Duong Nguyen Dao, Thanh Xuan Nguyen)

Link paper: https://aclanthology.org/2025.vlsp-1.37/

Presentation: https://www.youtube.com/watch?v=-klcvd7Ak8Q

---

🚀 I. Approach

Please refer to our paper for more details.

Our approach combines retrieval-based fine-tuning of large language models (Qwen2.5 and Qwen3) to enhance temporal reasoning and duration understanding in question answering tasks.

---

🛠 II. Training

Training Steps

Training is conducted using these scripts, each corresponding to a specific configuration:

• train_qwen2.5.py — Fine-tune Qwen2.5, output LoRA weights saved in saved_dir/.
• train_qwen3_prompt2.py — Fine-tune Qwen3 using an List output format prompt, weights saved in saved_dir_qwen3_prompt2/.
• train_qwen3_prompt2_aug.py — Fine-tune Qwen3 with cross-lingual data augmentation from UDST-DurationQA, weights saved in saved_dir_qwen3_prompt2_aug/.

Each model is trained independently.

---

⚙️ III. Inference & Submission

Step 1: Retrieval

Run retrieval scripts to compute similarity scores:

• retrieval.py → for public test
• retrieval_private.py → for private test

These scripts produce:

• similarities.npy, similarities_private.npy — Similarity matrices between test and training samples.
• results_retrieval.json, results_retrieval_private.json — Retrieved few-shot examples.

---

Step 2: Inference

Perform inference using five configurations:

Script	Description
infer_qwen2.5.py	Inference with the fine-tuned Qwen2.5 model.
infer_qwen3_prompt2.py	Inference with the fine-tuned Qwen3 model.
infer_qwen3_retrieval_prompt2.py	Inference with fine-tuned Qwen3 and retrieved few-shot samples.
infer_qwen3_prompt2_aug.py	Inference with fine-tuned Qwen3 using UDST-DurationQA augmentation.
infer_qwen3_retrieval_prompt2_aug.py	Inference with fine-tuned Qwen3 using augmentation + retrieval-based few-shot samples.

Output files are saved under submit/inference_*.csv.

---

Step 3: Submission

• Run submit_*.py scripts to clean inference results and produce final submission files (results_*.txt). The team only stored results from the public test runs and did not save results from the private test.
• To submit on AIHub, zip the final result:

  zip -r results.zip results.txt
  

Citation

If you find our work helpful, feel free to give us a cite.

@inproceedings{dao-nguyen-2025-retrieval,
    title = "Retrieval-Guided Fine-tuning for {V}ietnamese Event Duration Question Answering",
    author = "Dao, Duong Nguyen  and
      Nguyen, Thanh Xuan",
    editor = "Mai, Luong Chi  and
      Huyen, Nguyen Thi Minh  and
      Trang, Nguyen Thi Thu",
    booktitle = "Proceedings of the 11th International Workshop on Vietnamese Language and Speech Processing",
    month = oct,
    year = "2025",
    address = "Hanoi, Vietnam",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.vlsp-1.37/",
    pages = "306--315"
}