aadimodeldataset / README.md
Viharikvs's picture
Add dataset card
575a346 verified
|
Raw
History Blame Contribute Delete
3.94 kB
---
license: other
license_name: proprietary-urbankisaan
license_link: LICENSE
pretty_name: Aadi Training Data (plant genomics)
viewer: false
tags:
- biology
- genomics
- plant-genomics
- dna
- rna-seq
- atac-seq
- foundation-model
---
# Aadi Training Data
Raw and processed genomics data used to train **[Aadi](https://huggingface.co/Viharikvs/aadimodel)**,
a 404M-parameter plant-DNA foundation model by [UrbanKisaan Inc.](mailto:legal@urbankisaan.com)
This repository holds the inputs behind two stages: self-supervised pretraining of the shared
frozen trunk on 49 plant genomes, and supervised RNA-seq + ATAC coverage training for maize and
Arabidopsis.
> The dataset viewer is disabled on purpose. This repo contains raw sequencing reads, alignments,
> genome FASTA and coverage tracks (`.bam` / `.fastq.gz` / `.fa` / `.bw`), which are not tabular
> formats the viewer can render. Browse the files under **Files and versions** instead.
## Contents
| Component | Size | Files | What it is |
|---|---|---|---|
| `ssl_corpus/` | 70 GB | 49 genomes | Genome FASTA for self-supervised trunk pretraining (Ensembl Plants r56; AgroNT breadth corpus + one extra) |
| `maize_b73_v5/` | 393 GB | ~226 | Maize (B73 NAM 5.0) RNA-seq + ATAC: raw reads, alignments, coverage, indices |
| `arabidopsis_tair10/` | 173 GB | ~424 | Arabidopsis (TAIR10) RNA-seq + ATAC: raw reads, alignments, coverage, indices |
| **Total** | **~686 GB** | **~754** | |
## Directory layout
```
.
├── ssl_corpus/ 49 plant-genome FASTAs + .fai + agront_corpus_species.txt
│ ├── zea_mays.fa, oryza_sativa.fa, triticum_aestivum.fa, ...
│ └── agront_corpus_species.txt the species manifest
├── maize_b73_v5/
│ ├── fastq/ raw paired-end reads (.fastq.gz)
│ ├── rna_bam/ STAR RNA-seq alignments (*_Aligned.sortedByCoord.out.bam + .bai)
│ ├── atac_bam/ bowtie2 ATAC alignments (atac_*.bam + .bai)
│ ├── bigwig/ per-base CPM coverage tracks (.bw)
│ ├── indices/ STAR / bowtie2 genome indices
│ ├── reference/ Zm-B73-REFERENCE-NAM-5.0 genome (.fa) + annotation (.gtf/.gff3)
│ ├── metadata/ sample manifests
│ └── logs/ alignment logs
└── arabidopsis_tair10/ same subfolder structure, TAIR10
```
## Provenance
- **SSL corpus:** 49 plant genomes from Ensembl Plants release 56 — the AgroNT 48-species breadth
set plus one additional genome. Species manifest: `ssl_corpus/agront_corpus_species.txt`.
- **Reference genomes:**
- Maize — `Zea_mays.Zm-B73-REFERENCE-NAM-5.0`, Ensembl Plants r61 (FASTA + GTF + GFF3).
- Arabidopsis — `Arabidopsis_thaliana.TAIR10`, Ensembl Plants r56 (FASTA + GTF).
- **Reads:** public RNA-seq and ATAC-seq runs from ENA/SRA (accessions preserved in filenames,
e.g. RNA `ERR2273051`, ATAC `SRR4000469`; maize ATAC `SRR28817197`, etc.). RNA aligned with
STAR, ATAC with bowtie2; per-base coverage computed as CPM and stored as bigWig.
The full acquisition and preprocessing pipeline is scripted in the Aadi repo under
`scripts/data_pipeline/` (`fetch_agront_corpus.sh`, `maize_align_and_bigwig.py`).
## Relationship to the model
The frozen trunk and per-task adapters trained on this data live in
[`Viharikvs/aadimodel`](https://huggingface.co/Viharikvs/aadimodel). `ssl_corpus/` produced the
shared 404M trunk; the maize and Arabidopsis coverage data produced the supervised heads and the
cross-species transfer results.
## License
Proprietary — Copyright (c) 2026 UrbanKisaan Inc. All rights reserved. The underlying genome
assemblies and sequencing runs are publicly available from Ensembl Plants and ENA/SRA under their
respective terms; the curation, processing, and assembly of this collection are proprietary.