Simple Contrastive Splice Site Prediction Kevin Brar Reader's Note Paper's name in repo is SimCSP.pdf EXPERIMENTS: 1: relationship between f1/sccs in pretrain and fine-tune and its affects on F1 score uses the last layer's [CLS] token uses SpliceBERT-human, evaluated using f1/sccs on spliceator and NMI on hg19 could be improved by using SpliceBERT and NMI on spliceator 2 - hyperparameter search over pre-train w/ SCCS validation 3: experiment with chopping off 1 layers transformer and doing CL + fine-tune on that 4: experiment with chopping off 2 transformer layers and doing CL + fine-tune on that 5: freeze the transformer layers and only fine-tune the classifier this will tell you if the internal representation is being improved METRICS: SCCS : spearman correlation w/ cosine similarity NMI : normalized mutual information F1 : linear combination of precision and recall EXPERIMENTAL PIPELINE: pretrain.sh finetune.sh