Practical overview
When it fits
Use it when you have the Google QUEST competition data and want to study or reproduce the documented model-training and evaluation workflow.
Who it is for
- Researchers and practitioners seeking to reproduce the repository’s Google QUEST competition model workflow.
- Developers studying transformer-based question-and-answer labeling, model combinations or cross-validation analysis.
Documented uses
- Train and fine-tune the siamese RoBERTa, BERT and XLNet configurations and the double ALBERT configuration.
- Explore weighted ensembling and post-processing combinations with the stacking notebook.
- Review cross-validation scores under different data and target-handling settings.
Requirements & limits
The repository specifies an MIT license. Running the training workflow requires competition data, the listed Python packages and CUDA-capable GPU hardware.
- Training is documented for CUDA-enabled GPU hardware, with a single RTX 2080 Ti and about 10 GB GPU memory given as the default setup.
- The README suggests batch size 1 and gradient accumulation 8 as adjustments for CUDA out-of-memory errors.
- Using a different labeled Q&A dataset requires changes to data loading and preprocessing; compatibility with other formats is not demonstrated.
- The notebooks support analysis of ensembling, post-processing and cross-validation rather than providing a hosted labeling interface.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Reproduce the four competition model workflows
Train and fine-tune the four transformer configurations described in the repository.
Before you start
- Google QUEST competition data placed in a directory named data.
- Python packages listed in requirements.txt installed.
- CUDA-capable GPU; the README describes a single RTX 2080 Ti with about 10 GB of GPU memory as the default setup.
Steps
For each model name—siamese_roberta, siamese_bert, siamese_xlnet and double_albert—run the corresponding the command below command.
python train.py -model_name=After training each model, run the command below with the same model name.
python finetune.py -model_name=If CUDA out-of-memory errors occur, set batch size to 1 and gradient accumulation to 8 inside train.py, as suggested by the README.
Expected outcome
The documented training and fine-tuning stages are run for all four model configurations.
Guide limitations
- The recipe depends on the Google QUEST competition data and the packages in requirements.txt.
- The README gives a default GPU setup but does not establish requirements for other hardware configurations.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- robinniesert/kaggle-google-quest · Official homepage title · retained in its source language