---
tags:
  - DNA
license: mit
---

## MiniDNA model

This is a distilled version of [DNABERT](https://github.com/jerryji1993/DNABERT) by using MiniLM technique. It has a BERT architecture with 6 layers and 768 hidden units, pre-trained on 6-mer DNA sequences. For more details on the pre-training scheme and methods, please check the original thesis report _[link to be added]_.


## How to Use  

The model can be used to fine-tune on a downstream genomic task, e.g. promoter identification.

```python
import torch
from transformers import BertForSequenceClassification
model = BertForSequenceClassification.from_pretrained('Peltarion/dnabert-minilm')
```

More details on how to fine-tune the model, dataset and additional source codes are available on [github.com/joanaapa/Distillation-DNABERT-Promoter](https://github.com/joanaapa/Distillation-DNABERT-Promoter).