---
configs:
- config_name: binary
  data_files:
  - split: train
    path: binary/train-*
- config_name: clinvar_levels
  data_files:
  - split: train
    path: clinvar_levels/train-*
- config_name: pathogenic_pairwise
  data_files:
  - split: train
    path: pathogenic_pairwise/train-*
- config_name: all_tasks
  data_files:
  - split: train
    path: all_tasks/train-*
---

# VEPQA

VEPQA is a supervised fine-tuning dataset for variant-effect-prediction style question answering over DNA sequence context. It is built from raw ClinVar variant records and UCSC hg38 reference sequence windows, with deterministic gold labels from ClinVar.

## Configs

- `binary`: binary clinical-significance classification.
- `clinvar_levels`: finer-grained ClinVar clinical-significance classification.
- `pathogenic_pairwise`: reference-vs-alternate sequence preference for pathogenic or likely pathogenic variants.
- `all_tasks`: concatenation of the three task configs.

## Row Format

Each row includes:

- `question`: the task prompt body before the reusable answer instruction/options suffix.
- `prompt_suffix`: the reusable answer instruction/options suffix.
- `prompt`: the full user-facing prompt, equal to `question + "\n\n" + prompt_suffix`.
- `options`: allowed answer labels for the task.
- `answer_index`: index of the correct option.
- `answer_text`: deterministic plain-text target.
- `messages`: two-turn SFT messages, with the plain answer in the assistant message.
- `gold_answer_json`: JSON-form audit target, not the primary SFT target.
- audit metadata: ClinVar IDs, genomic position, review status, sequence hashes, and benchmark exclusion status.

For evaluation, use `prompt` as model input and compare to `answer_text` or `answer_index`. Use `question` and `prompt_suffix` when you need to inspect or vary the prompt template. The `messages` column is intended for SFT and includes the answer.

## Provenance

Source variant table:

- ClinVar `variant_summary.txt.gz`.

Reference sequence:

- UCSC hg38 per-chromosome FASTA files.

Carbon eval exclusion manifest:

- BRCA2 VEP: `HuggingFaceBio/brca2-vep`
- TraitGym Mendelian VEP: `HuggingFaceBio/traitgym`, Mendelian VEP parquet
- Carbon ClinVar VEP: `HuggingFaceBio/clinvar-vep-final`, coding and non-coding parquets

The dataset was generated from raw ClinVar records, not from Carbon benchmark rows.

Dataset-wide provenance summary:

```json
{
  "source": {
    "clinvar_variant_summary": 6000
  },
  "assembly": {
    "GRCh38": 6000
  },
  "origin": {
    "germline": 6000
  },
  "variant_type": {
    "single nucleotide variant": 6000
  },
  "window_bp": {
    "8192": 6000
  },
  "variant_window_placement": {
    "final": 6000
  },
  "variant_index_0based": {
    "8191": 6000
  },
  "variant_position_1based": {
    "8192": 6000
  },
  "benchmark_exclusion_status": {
    "checked": 6000
  }
}
```

Every prompt sequence is 8,192 bp long and is wrapped in closed DNA tags, for example `<dna>...</dna>`. The variant is placed at the final base of the sequence context: `variant_index_0based = 8191`, equivalent to `variant_position_1based = 8192`.

The `all_tasks` config has 6,000 task rows but 4,000 unique variant keys. The `binary` and `clinvar_levels` configs are two views of the same 2,000 clinical-classification variants; `pathogenic_pairwise` contributes 2,000 additional pathogenic or likely pathogenic variants.

Unique variant chromosome distribution across `all_tasks`:

```json
{
  "1": 375,
  "2": 395,
  "3": 193,
  "4": 100,
  "5": 180,
  "6": 150,
  "7": 209,
  "8": 109,
  "9": 164,
  "10": 128,
  "11": 261,
  "12": 209,
  "13": 69,
  "14": 92,
  "15": 139,
  "16": 205,
  "17": 344,
  "18": 49,
  "19": 213,
  "20": 81,
  "21": 45,
  "22": 72,
  "X": 218
}
```

Task-row chromosome distribution in `all_tasks`:

```json
{
  "1": 572,
  "2": 571,
  "3": 283,
  "4": 160,
  "5": 271,
  "6": 216,
  "7": 316,
  "8": 169,
  "9": 258,
  "10": 190,
  "11": 386,
  "12": 310,
  "13": 103,
  "14": 146,
  "15": 198,
  "16": 324,
  "17": 511,
  "18": 78,
  "19": 321,
  "20": 129,
  "21": 74,
  "22": 112,
  "X": 302
}
```

## ClinVar Filtering

Rows were retained from ClinVar only if they met all of the following criteria:

- Assembly: `GRCh38`.
- Variant type: single nucleotide variant.
- Origin: germline.
- Chromosome: autosomes or `X`.
- Clean VCF alleles: one-base `A/C/G/T` reference and alternate alleles.
- Clinical significance in one of:
  - `Benign`
  - `Likely benign`
  - `Benign/Likely benign`
  - `Likely pathogenic`
  - `Pathogenic`
  - `Pathogenic/Likely pathogenic`
- Review status in the strict set:
  - `practice guideline`
  - `reviewed by expert panel`
  - `criteria provided, multiple submitters, no conflicts`

The larger `criteria provided, single submitter` class was intentionally not used.

## Eval-Overlap Exclusion

To avoid overlap with Carbon VEP evaluations, a manifest was built from the Carbon BRCA2, TraitGym Mendelian, and ClinVar VEP evaluation datasets. Candidate ClinVar SFT rows were excluded if either:

- the exact variant key `(assembly, chrom, pos, ref, alt)` appeared in the Carbon eval manifest; or
- the candidate position fell inside an expanded eval interval in the same assembly.

The interval buffers used in the manifest were:

- BRCA2: 8,192 bp buffer around each eval variant.
- TraitGym Mendelian: 8,192 bp buffer around each eval variant.
- Carbon ClinVar VEP: 24,000 bp buffer around each eval variant.

Carbon's BRCA2 eval rows are represented in `hg19`, while this dataset is `GRCh38`. To avoid relying on liftover during this build, the dataset additionally hard-excludes a conservative GRCh38 BRCA2 locus interval:

```text
chr13:32,300,000-32,420,000
```

After these gates, the sampled dataset has:

```text
benchmark_exclusion_status: checked
excluded_from_benchmark_eval: true
```

## Task Design

### `binary`

The prompt states the mutation location in natural language and shows an 8,192 bp DNA sequence window containing the alternate allele at the final base of the context. For example, it tells the model that the variant is at position 8,192 using 1-based indexing, and that the reference allele is changed to the alternate allele at that position. The model is asked to answer with exactly one option:

```text
benign or likely benign
pathogenic or likely pathogenic
```

This config collapses ClinVar benign/likely-benign labels and pathogenic/likely-pathogenic labels into two classes. The collapsed natural-text label is retained in `label_binary`, and the model-facing answer text uses the same format.

### `clinvar_levels`

This task uses the same sequence prompt style as `binary`, including the natural-language mutation location sentence, but preserves the retained raw ClinVar label. Allowed answers are:

```text
Benign
Likely benign
Benign/Likely benign
Likely pathogenic
Pathogenic/Likely pathogenic
Pathogenic
```

### `pathogenic_pairwise`

This task compares the reference and alternate sequence for a single ClinVar pathogenic or likely pathogenic SNV. Sequence A/B order is randomized. Because the alternate allele is labeled pathogenic or likely pathogenic, the reference sequence is treated as more consistent with normal biological function.

The model is asked to answer with exactly one option:

```text
Sequence A
Sequence B
```

This task is not a benign-vs-pathogenic matched-pair task. Benign or likely benign alternate alleles are not used for this reference-vs-alternate target, because a benign alternate allele is not necessarily less functional than the reference.

## Data Distribution

Config sizes:

```json
{
  "binary": 2000,
  "clinvar_levels": 2000,
  "pathogenic_pairwise": 2000,
  "all_tasks": 6000
}
```

Clinical-classification raw-label distribution, used by both `binary` and `clinvar_levels`:

```json
{
  "Benign": 333,
  "Benign/Likely benign": 334,
  "Likely benign": 333,
  "Likely pathogenic": 333,
  "Pathogenic": 334,
  "Pathogenic/Likely pathogenic": 333
}
```

`pathogenic_pairwise` raw-label distribution:

```json
{
  "Likely pathogenic": 659,
  "Pathogenic": 670,
  "Pathogenic/Likely pathogenic": 671
}
```

Review-status distribution over the 4,000 unique prompt rows:

```json
{
  "criteria provided, multiple submitters, no conflicts": 3651,
  "reviewed by expert panel": 348,
  "practice guideline": 1
}
```

Build-stage counts:

```json
{
  "raw_rows_scanned": 8907730,
  "strict_clinvar_candidates": 333140,
  "after_carbon_eval_manifest_gate": 333140,
  "excluded_by_grch38_brca2_locus": 3818,
  "after_grch38_brca2_locus_exclusion": 329322,
  "sampled_prompt_rows": 4000
}
```

## Limitations

- Labels are ClinVar clinical-significance labels, not direct experimental functional measurements.
- `review_status` and submitter counts are included as metadata but are not shown in the prompt; the model is not trained to infer evidence strength.
- The current release has only a `train` split. A larger training run should introduce region-disjoint validation data with an interval buffer.
- The pairwise task is reference-vs-pathogenic-alternate only. A benign-vs-pathogenic matched-pair task would require separate matching logic.