daikichiba9511/atma17

AtmaCup 17

competition page: https://www.guruguru.science/competitions/24

Description

• 12 th place solution for AtmaCup 17
• Task: predict the item's recommendation from data including review text.
• solution
  • please see solution.md.
  • Discussion post version (Ja): https://www.guruguru.science/competitions/24/discussions/2a97a005-de13-4e2b-87f4-1b2b35d8011e/

Usage

1. Clone this repository

2. Download datesets from the competition page

3. unzip the datasets and put them in the input directory

4. Docker build and Run

   docker compose up -d
   # to interact with the container
   docker compose local-dev exec bash

5. run the following command

   make setup
   python -m src.exp.exp002.train && python -m src.exp.exp002.test
   python -m src.exp.exp004.train && python -m src.exp.exp004.test
   python -m src.exp.exp006.train && python -m src.exp.exp006.test
   python -m ensemble

References

[1] takaito, [atmaCup17] Tutorial Notebook① DeBERTa Large Modelの学習 (少し手を加えるとLB: 0.9718), https://www.guruguru.science/competitions/24/discussions/4aee3312-ba40-4934-9035-22c7d2b51c09/

[2] nishimoto, 2023-24年のKaggleコンペから学ぶ、NLPコンペの精度の上げ方, https://zenn.dev/nishimoto/articles/974f2a445f9d74

[3] AI SHIFT, 【AI Shift/Kaggle Advent Calendar 2022】Kaggleで学んだBERTをfine-tuningする際のTips④〜Adversarial Training編〜, https://www.ai-shift.co.jp/techblog/2985