| --- |
| license: apache-2.0 |
| task_categories: |
| - text-generation |
| tags: |
| - biology |
| - genomics |
| - dna |
| - benchmark |
| configs: |
| - config_name: tata |
| data_files: |
| - split: train |
| path: tata/data.parquet |
| - config_name: synonymous_codons |
| data_files: |
| - split: train |
| path: synonymous_codons/data.parquet |
| --- |
| |
| # Carbon Perturbation Bench |
|
|
| Two **sequence-level perturbation tasks** for evaluating DNA foundation models: |
|
|
| - **`tata`** (5,000 rows): the TATA-box motif inside a real promoter is |
| disrupted with random nucleotide substitutions. Probes whether the model |
| has internalised eukaryotic promoter architecture (it should assign higher |
| log-likelihood to the intact promoter than the perturbed one). |
| - **`synonymous_codons`** (5,000 rows): codons in a real CDS are replaced |
| with synonyms encoding the same amino acid. Probes whether the model has |
| learned codon-usage bias (it should prefer native codon usage over the |
| synonymous variant). |
| |
| Both subsets share the same schema: |
| |
| | column | description | |
| |---|---| |
| | `chr`, `start`, `end`, `strand` | hg38 locus | |
| | `length` | sequence length in bp | |
| | `original_sequence` | the real (unperturbed) sequence — positive | |
| | `sequence` | the perturbed sequence — negative control | |
| |
| ## Usage |
| |
| ```python |
| from datasets import load_dataset |
| |
| tata = load_dataset("hf-carbon/carbon-perturbation-bench", "tata", split="train") |
| syn = load_dataset("hf-carbon/carbon-perturbation-bench", "synonymous_codons", split="train") |
| ``` |
| |
| ## Eval recipe |
| |
| Pairwise likelihood discrimination: score `LL(original_sequence)` vs |
| `LL(sequence)` under the model, and report `mean(LL(real) >= LL(perturbed))`. |
| A ready-to-run scorer for Carbon, GENERator, and Evo2 lives at |
| [`evaluation/perturbation_tasks.py`](https://github.com/huggingface/carbon) |
| in the Carbon release repo. |
| |
