Table 1 Impact of Max Sequence Length on CORAL. Training on markdown data in addition to code significantly increases performance independent of maximum sequence length
From: CORAL: COde RepresentAtion learning with weakly-supervised transformers for analyzing data analysis
Model | Max Sequence Length | ||
|---|---|---|---|
80 | 120 | 160 | |
CORAL | 59.9 | 64.2 | 72.2 |
CORAL(No Markdown) | 54.4 | 57.0 | 60.2 |