deftriclm

# DefTriCLM task-specific models and pretraining-corpus mining results This dataset repository contains the released outputs of the DefTriCLM experiments for three gene-level prediction tasks: - BGC-associated gene prediction - TXSS-associated gene prediction - defense-gene prediction ## Repository layout ```text models/ BGC/ best.pt, last.pt, args.json TXSS/ best.pt, last.pt, args.json defense/ best.pt, last.pt, args.json mining/ BGC/ predictions and embeddings over the pretraining corpus TXSS/ predictions and embeddings over the pretraining corpus defense/ predictions and embeddings over the pretraining corpus manifests/ mining/ mining completion and integrity metadata training/ fine-tuning completion metadata splits/ task_ttv_final_20260810/ deterministic genome-holdout TTV IDs and audits ``` ## TTV split policy BGC and TXSS fine-tuning use independent, deterministic genome-level train/validation/test holdouts. Each genome belongs to exactly one split, so gene IDs from the same genome cannot cross splits. Cluster overlap is audited in the included split artifacts but is not used as a hard grouping constraint. ## Checkpoints `best.pt` is selected by validation gene-level average precision. `last.pt` stores the final training state when available. The corresponding `args.json` records the complete run configuration. Checkpoints are PyTorch files and must only be loaded in a trusted environment. ## Mining outputs The three `mining/` directories contain predictions produced by the task-specific checkpoints over the complete prepared DefTriCLM pretraining shard collection. Embeddings were retained together with the predictions. Use the included manifests to enumerate chunks and verify completeness instead of inferring completeness from directory listings alone. ## Notes - This release contains derived predictions, embeddings, experiment metadata, split ID lists, and model checkpoints. It does not duplicate the raw source biological sequence datasets. - Review the original source-dataset licenses before redistributing or deriving additional public datasets. - Large files are uploaded with the ModelScope resumable upload client.