---
library_name: transformers
license: mit
tags:
- genomics
- dna
- opticaldna
- vision-language
- custom-code
datasets:
- hxxiang/dna_benchmarks
---

# OpticalDNA-HG38-2048

ICML 2026 OpticalDNA checkpoint pretrained on HG38 with 2048-base rendered DNA documents. Released checkpoint: **step 190,000**.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "hxxiang/opticaldna-hg38-2048"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype="auto",
).cuda().eval()

# Visual features only: the language Decoder is not executed.
features = model.extract_features("page.png", pooling="mean", to_cpu=True)

# Decoder inference: short T1 / Free OCR.
prompt = model.build_prompt("t1_full_ocr", length="short")
text = model.generate_document(
    tokenizer,
    "page.png",
    prompt=prompt,
    max_new_tokens=256,
)

# Prompt-conditioned Decoder hidden states.
decoder_features = model.extract_decoder_features(
    tokenizer,
    "page.png",
    prompt=prompt,
    pooling="mean",
    to_cpu=True,
)
```

For multi-page input, pass pages in reading order, e.g. `["page1.png", "page2.png"]`.

Code and environment: [OpticalDNA GitHub](https://github.com/HongxinXiang/OpticalDNA). Paper: *Rethinking Genomic Modeling Through Optical Character Recognition*, ICML 2026.