Analyze training results

The training report allows you to assess the models performance, so as to proceed to the next step, or modify the dataset or training input and re-run. The most important metric is the Spearman rank correlation.

Spearman rank correlation

Spearman measures how well the model's predicted rankings match the actual rankings in your data: +1 would mean perfect agreement i.e. every variant ranked higher really measured better, -1 would mean the reverse, and 0 would mean no correlation.

The correlation is measured on held-out data the model never saw in training, so it reflects the ability to predict new variants, not memorization.

A low Spearman in round one can happen if there is little data available (less than a couple of hundreds). It will improve with each round. Depending on data resolution and quantity, a minimum Spearman rank of about 0.4 (with a significant p-value, < 0.05) is sufficient to expect good results.

Poor model performance may have multiple causes: data quantity and/or quality, weak sequence-property signal, etc. You may want to review your data and/or model configurations, and re-run. See the Troubleshooting section for more details.