FishCaduceus-28L-1024-FishGUE

FishCaduceus-28L-1024-FishGUE is a collection of 17 task-specific models obtained by fine-tuning the pretrained FishCaduceus-28L-1024 DNA language model on downstream fish genomic function prediction tasks from Fish_GUE.

Model description

This repository contains one FishCaduceus backbone family and one fine-tuned sequence-classification model for each included task. The original task directory names are retained so that the repository can be used with subfolder-based loading.

Base model

The base model is FishCaduceus/FishCaduceus-28L-1024.

Included downstream tasks

Task directory	Task description
CTCF_TF	CTCF transcription factor binding site prediction
H3k27ac	H3K27ac histone modification prediction
H3K27me3	H3K27me3 histone modification prediction
H3K4me1	H3K4me1 histone modification prediction
H3K4me3	H3K4me3 histone modification prediction
H3K9me3	H3K9me3 histone modification prediction
Pou5f1_TF	Pou5f1 transcription factor binding site prediction
Sox2_TF	Sox2 transcription factor binding site prediction
prom_1k_all	Promoter prediction using 1-kb sequences
prom_1k_notata	Promoter prediction using 1-kb sequences for the non-TATA promoter subset
prom_1k_tata	Promoter prediction using 1-kb sequences for the TATA promoter subset
prom_300_all	Promoter prediction using 300-bp sequences
prom_300_notata	Promoter prediction using 300-bp sequences for the non-TATA promoter subset
prom_300_tata	Promoter prediction using 300-bp sequences for the TATA promoter subset
splice_acceptor	Splice acceptor site prediction
splice_donor	Splice donor site prediction
splice_all	Splice site classification

Repository structure

The repository contains 17 task directories. Each task retains its original name and contains a configuration, model weight file, and tokenizer files where present.

Training data

These models were fine-tuned on the corresponding downstream genomic function prediction tasks from Fish_GUE. No task sample counts are asserted here.

Intended uses

The models are intended for research on fish genomic sequence representation and task-specific genomic function prediction.

Limitations

The task models reflect the training data and task definitions used by Fish_GUE. Transfer to other species, assemblies, sequence lengths, or label definitions should be validated independently.

How to use

The audited task configurations declare CaduceusForSequenceClassification, model_type: caduceus, binary or three-way label mappings through id2label/label2id, and an auto_map for Caduceus classes. Loading with Transformers is expected to require trust_remote_code=True and an environment that provides the referenced Caduceus implementation. The task directories do not contain custom Python implementation files, so standalone execution from this repository alone is not guaranteed.

Files in this repository

Intermediate checkpoints, test metrics, test predictions, and training argument files were deliberately excluded. Model weights, configurations, and tokenizer files needed by the original task exports were retained.

Citation

The FishCaduceus manuscript is in preparation. Citation information will be added after publication.

Acknowledgements

FishCaduceus was developed for research on fish genomes at the Institute of Hydrobiology, Chinese Academy of Sciences.

Contact

Xiao-Qin Xia
Institute of Hydrobiology, Chinese Academy of Sciences
Email: xqxia@ihb.ac.cn