# Genomic-NIAH: needle-in-a-haystack long-context retrieval benchmark for DNA language models

## Facts

- Provider: Hugging Face (HuggingFaceBio, Carbon team)
- License: Apache-2.0 (https://huggingface.co/datasets/HuggingFaceBio/genomic-niah)
- Size: 28 files
- Format: PNG, Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/HuggingFaceBio/genomic-niah
- Access: public, readable without an account
- Category: Genomics & Bioinformatics
- Source: https://huggingface.co/datasets/HuggingFaceBio/genomic-niah
- Page: https://scigantic.com/archive/7413e9e2-c9ea-5349-9d16-957fec78ecd6
- Keywords: Genomics & Bioinformatics, benchmark, PNG, Parquet, Hugging Face

## About

A long-context retrieval benchmark for genomic language models from Hugging Face, built for the Carbon DNA models. Each example plants a random 24 bp KEY and 24 bp VALUE pair inside a real eukaryotic genome haystack and asks the model to return VALUE given the haystack followed by KEY, so success requires actual long-range retrieval. Four tasks (plain retrieval and three near-duplicate variants with eight distractor keys differing by 4, 2 or 1 bp, inspired by RULER multi-key) cross six context lengths from 4k to 128k 6-mer tokens (about 24 kbp to 786 kbp), giving 24 configurations of 500 examples each (12,000 in total), stratified over five needle depths (10 to 90%) and four kingdoms (animalia, fungi, plantae, protista). Haystacks come from the OpenGenome2 NCBI eukaryotic test and validation splits, using contiguous ACGT runs up to 16k tokens and stitched within-record runs beyond. Data are Parquet under data/, one config per task and length (e.g. niah_16k, niah_neardup_d1_64k), loadable with Hugging Face datasets; about 4 GB. Released under Apache 2.0; cite the Carbon paper (Ben Allal et al., 2026).

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("HuggingFaceBio/genomic-niah")
print(ds)
```

## Data structure

File types in sampled files: PNG, Parquet

Example file: `data/niah_128k/test-00000-of-00001.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/7413e9e2-c9ea-5349-9d16-957fec78ecd6
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/7413e9e2-c9ea-5349-9d16-957fec78ecd6
- Over MCP: `get_schema_card` with id `7413e9e2-c9ea-5349-9d16-957fec78ecd6` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Genomics & Bioinformatics datasets](https://scigantic.com/topics/genomics-and-bioinformatics)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [PNG datasets](https://scigantic.com/topics/png)
