AmelieSchreiber/esm2_t33_650M_qlora_binding_16M

Updated Nov 21, 2023 • 7

Note A QLoRA trained on ~16M protein sequences with binding site annotations from UniProt.

AmelieSchreiber/esm2_t33_650M_qlora_binding_12M

Updated Nov 1, 2023

Note A QLoRA trained on ~12M protein sequences with binding site annotations from UniProt.

AmelieSchreiber/esm2_t6_8m_qlora_binding_sites_v0

Updated Oct 4, 2023 • 1

Note While this model is not overfit, it still shows more signs of overfitting than a model that uses more QLoRA adapter layer. This model only uses adapters for the query, key, and value weight matrices.

AmelieSchreiber/esm2_t12_35M_qlora_binding_sites_v0

Updated Sep 29, 2023

Note This model only uses adapters for the query, key, and value weight matrices. Thus, it is more overfit than a model that uses more adapter layers.

AmelieSchreiber/esm2_t6_8m_qlora_binding_sites_v1

Updated Sep 30, 2023

Note Less overfitting occurs in this model due to more weight matrices being adapted with QLoRA.

AmelieSchreiber/esm2_t12_35M_qlora_binding_sites_v1

Updated Oct 6, 2023

Note Less overfitting occurs in this model due to more weight matrices being adapted with QLoRA.

Note This dataset is curated from UniProt. The test set was created by selecting entire families of proteins to separate out at random. The train/test split is approximately 80/20. All binding site and active site annotations were merged. All sequences longer than 1000 amino acids were split into non-overlapping chunks of 1000 residues or less.

Note This dataset is curated from UniProt. The test set was created by selecting entire families of proteins to separate out at random. The train/test split is approximately 80/20. All binding site and active site annotations were merged. All sequences longer than 1000 amino acids were split into non-overlapping chunks of 1000 residues or less.

Note This dataset is curated from UniProt. The test set was created by selecting entire families of proteins to separate out at random. The train/test split is approximately 80/20. All binding site and active site annotations were merged. All sequences longer than 1000 amino acids were split into non-overlapping chunks of 1000 residues or less.