---
license: apache-2.0
tags:
- biology
- genomics
- dna
- genezip
- pretrained
---

# GeneZip-Intergenic

GeneZip pretrained DNA language model, published under its Camera Ready name.
This repository contains pretrained weights, not a downstream task checkpoint.

- Pretraining context: 12,800 bp.
- Base compression target: 32 bp/token (not the measured compression ratio).
- Source: [jzshared/GeneZip-AntiBioPrior-BPT32-12.8K](https://huggingface.co/jzshared/GeneZip-AntiBioPrior-BPT32-12.8K/tree/3b9c82577ee4159ef6ab3421f4bca45ab2fee5f8/checkpoint-30600).
- Weight SHA-256: `a35a52717e68e26abae507b5a9dc547f60ae6de19e0aeb56345bd0c3ce525d89`.
- Weight bytes are unchanged from this pinned source. The Hydra configuration is
  reduced to architecture, routing and scientific metadata.

## Code and usage

Use the GeneZip Camera Ready code from [AndyJZhao/GeneZip](https://github.com/AndyJZhao/GeneZip).
The CR code update is being prepared separately; these weights have been checked
with its local release candidate. This is a custom GeneZip checkpoint, not a
Transformers AutoModel package.

```python
from huggingface_hub import snapshot_download
checkpoint = snapshot_download("andyjzhao/GeneZip-Intergenic")
```

The release code accepts `ckpt=andyjzhao/GeneZip-Intergenic` for downstream training. For an explicit
local download, pass the returned directory as `ckpt`. The pretraining evaluator
uses `pretrained_ckpt` with the matching model/context configuration.

`hydra_cfg.yaml` records the exact architecture and routing bounds.
`provenance.json` records the immutable source revision and source weight hash.
The repository contains no downstream fine-tuned weights or optimizer state.

## Intended use and limitations

Research on genomic sequence modeling and the GeneZip DNALongBench recipes.
Short smoke checks establish loading and execution; they are not new measurements
of full benchmark performance. Refer to the code's result mapping for adopted runs.