Upload README.md

README.md

@@ -1,149 +1,30 @@
-<div align="center">
-
 # FishCaduceus
 
-### Single-nucleotide DNA language models and benchmarks for fish genomics
-
-**Open models, datasets and evaluation resources for cross-species functional annotation, evolutionary constraint analysis and genetic variant interpretation in fish genomes.**
-
-</div>
-
----
-
-## About FishCaduceus
-
-**FishCaduceus** is a family of fish-specific DNA language models developed for cyprinid and comparative fish genomics.
-
-The models were pretrained on reference genomes from seven major cyprinid aquaculture fishes and zebrafish. FishCaduceus operates at single-nucleotide resolution and combines the Caduceus and Mamba architectures with bidirectional sequence modeling and reverse-complement equivariance.
-
-FishCaduceus is designed to learn transferable functional and evolutionary information directly from unlabeled genomic sequences. The project currently supports research on:
-
-- masked nucleotide prediction and genomic representation learning
-- cross-species gene-annotation transfer
-- translation initiation and termination site prediction
-- splice donor and acceptor site prediction
-- evolutionary constraint prediction
-- zero-shot mutation effect scoring
-- prioritization of candidate functional variants
-
-In the accompanying study, FishCaduceus representations trained with zebrafish annotations generalized to six additional fish species across four functional-site prediction tasks. The best model achieved an average cross-species AUPRC of **0.957**. Evolutionary constraint benchmarks were further constructed from a whole-genome alignment of 26 fish species.
-
----
-
-## Models
-
-| Model | Parameters | Context length | Description |
-|---|---:|---:|---|
-| [FishCaduceus-20L-512](https://huggingface.co/FishCaduceus/FishCaduceus-20L-512) | 20.9M | 512 nt | Compact 20-layer FishCaduceus model |
-| [FishCaduceus-28L-512](https://huggingface.co/FishCaduceus/FishCaduceus-28L-512) | 112.1M | 512 nt | Larger 28-layer model with a 512-nt context |
-| [FishCaduceus-28L-1024](https://huggingface.co/FishCaduceus/FishCaduceus-28L-1024) | 112.1M | 1,024 nt | Larger 28-layer model with an extended context |
-
-All models use:
-
-- single-nucleotide tokenization
-- masked language modeling
-- bidirectional sequence processing
-- reverse-complement-aware representations
-- custom Hugging Face Transformers code
-
----
-
-## Pretraining datasets
-
-| Dataset | Sequences | Sequence length | Description |
-|---|---:|---:|---|
-| [FishCaduceus Pretraining Dataset 512](https://huggingface.co/datasets/FishCaduceus/FishCaduceus-Pietraining-512) | 6,087,221 | 512 bp | Stratified genomic sequence corpus for 512-nt models |
-| [FishCaduceus Pretraining Dataset 1024](https://huggingface.co/datasets/FishCaduceus/FishCaduceus-Pietraining-1024) | 2,709,306 | 1,024 bp | Stratified genomic sequence corpus for the 1,024-nt model |
-
-The pretraining corpora were constructed from the following species:
-
-- *Hypophthalmichthys nobilis*
-- *Hypophthalmichthys molitrix*
-- *Mylopharyngodon piceus*
-- *Cyprinus carpio*
-- *Megalobrama amblycephala*
-- *Ctenopharyngodon idella*
-- *Carassius gibelio*
-- *Danio rerio*
-
-Species-aware and repeat-content-aware stratified sampling was used to reduce imbalances caused by differences in genome size and repeat composition.
-
----
+**FishCaduceus** is a family of single-nucleotide DNA language models developed for cyprinid and comparative fish genomics.
 
-## Benchmark datasets
+The project provides:
 
-### Functional site prediction
+- pretrained FishCaduceus models with 512-bp and 1,024-bp context lengths
+- pretraining datasets from seven major cyprinid aquaculture fishes and zebrafish
+- cross-species benchmarks for translation initiation, translation termination, splice donor and splice acceptor prediction
+- evolutionary constraint benchmarks derived from a 26-fish whole-genome alignment
 
-[**FishCaduceus Functional Site Benchmark**](https://huggingface.co/datasets/FishCaduceus/FishCaduceus-Functional-Site-Benchmark)
+FishCaduceus is designed for fish genome annotation, cross-species transfer, evolutionary constraint analysis and zero-shot variant prioritization.
 
-Sequence-based benchmarks for four gene-annotation tasks:
+## Resources
 
-- translation initiation sites
-- translation termination sites
-- splice donor sites
-- splice acceptor sites
-
-The benchmark contains fixed zebrafish training, validation and held-out test splits, together with independent test sets from six additional fish species. The benchmark is designed for direct cross-species evaluation without target-species retraining.
-
-### Evolutionary constraint prediction
-
-[**FishCaduceus Evolutionary Constraint Benchmark**](https://huggingface.co/datasets/FishCaduceus/FishCaduceus-Evolutionary-Constraint-Benchmark)
-
-Sequence-based benchmarks derived from a 26-fish Progressive Cactus whole-genome alignment.
-
-The repository includes:
-
-- grass carp training, validation and held-out test splits
-- cross-species whole-genome test sets
-- CDS-, intron- and intergenic-region test sets
-- highly conserved and low-conservation site labels
-
----
-
-## Recommended model
-
-For most applications, we recommend starting with:
-
-[**FishCaduceus-28L-1024**](https://huggingface.co/FishCaduceus/FishCaduceus-28L-1024)
-
-Use the 512-nt models when lower memory consumption or faster inference is preferred.
-
----
-
-## Intended use
-
-FishCaduceus resources are intended for research in:
-
-- fish genomics
-- aquaculture genomics
-- comparative genomics
-- genome annotation
-- DNA language modeling
-- evolutionary genomics
-- genetic variant prioritization
-
-Model predictions reflect statistical patterns learned from genomic sequences. They do not by themselves demonstrate biological function or causality and should be interpreted together with genetic, evolutionary and experimental evidence.
-
----
+- Models: [FishCaduceus organization](https://huggingface.co/FishCaduceus)
+- Functional-site benchmark: [FishCaduceus-Functional-Site-Benchmark](https://huggingface.co/datasets/FishCaduceus/FishCaduceus-Functional-Site-Benchmark)
+- Evolutionary-constraint benchmark: [FishCaduceus-Evolutionary-Constraint-Benchmark](https://huggingface.co/datasets/FishCaduceus/FishCaduceus-Evolutionary-Constraint-Benchmark)
 
 ## Manuscript
 
 **FishCaduceus: A DNA Language Model for Cyprinid Genomics at Single-Nucleotide Resolution**
 
-The manuscript is currently in preparation. Full citation information will be added after publication.
-
----
-
-## Development
-
-FishCaduceus was developed by researchers from the Institute of Hydrobiology, Chinese Academy of Sciences, together with collaborating institutions.
-
-The project builds on the Caduceus architecture and related open-source genomic sequence-modeling software.
-
----
+The manuscript is in preparation.
 
 ## Contact
 
-**Xiao-Qin Xia**  
+Xiao-Qin Xia  
 Institute of Hydrobiology, Chinese Academy of Sciences  
-Email: xqxia@ihb.ac.cn
+Email: xqxia@ihb.ac.cn