80/20: Contextual Training Effects on Visual Feature Learning

Experiment testing whether training context (isolated vs. bigram) changes how vision models learn Chinese characters. Demonstrates that the standard 80/20 train/test split is an experimental variable, not a neutral default.

Quick Start

pip install -r requirements.txt streamlit run app.py

Data

CASIA-HWDB (Liu et al., 2011) via HuggingFace Teklia/CASIA-HWDB2-line CalliBench (Luo et al., 2025) for transfer evaluation

Key Finding

Split ratio changes internal validation by 21+ percentage points. The 80/20 test/train default is an imputed, methodological decision that should be assessed per use case. Its companion Streamlit dashboard can be found here: https://tinyurl.com/8020ml