# OligoVigil benchmark reference splits

Release date: 2026-06-14  
Resource DOI: 10.5281/zenodo.20633779  
Website: https://oligovigil.pages.dev/

## Scope

The benchmark is a small, versioned reference split for reproducible evaluation on curator-verified therapeutic-oligonucleotide safety and off-target evidence. It is not a large-scale toxicity prediction corpus and should not be interpreted as a de novo risk model.

## Tasks

| task_name | target field | train | validation | test | Grade A | Grade B |
|---|---:|---:|---:|---:|---:|---:|
| toxicity_safety_v0_1 | toxicity endpoint category | 218 | 23 | 22 | 180 | 83 |
| offtarget_safety_v0_1 | off-target evidence type | 66 | 5 | 10 | 32 | 49 |

## Eligibility

Only release records with all of the following are eligible:

- curator-verified accept audit
- evidence grade A or B
- explicit evidence domain: `toxicity` or `offtarget`
- source and molecule/cohort fields sufficient to form a leakage group

Grade C records are released for browsing and citation but are excluded from reference splits.

## Split policy

Rows are grouped by `leakage_group`, defined as a source-paper by molecule/cohort group. No leakage group appears in more than one split within a task. The current public split file contains 344 rows and has zero leakage-group cross-split violations.

## Files

- `benchmark_reference_splits.csv`: fixed row-level train/validation/test assignments.
- `benchmark_task_cards.csv`: task definitions, label fields, eligibility rules and recommended metrics.
- `benchmark_baseline_results.csv`: deterministic sanity baselines.
- `evidence_release.csv`: full verified release table used to interpret benchmark rows.

## Baselines

Included baselines are deterministic floor estimates:

- training-set majority class
- modality-prior class
- evidence-grade-prior class
- target-prior class

These baselines are not proposed models. They are provided to make split reuse checkable and to expose trivial-prior performance before training stronger methods.

## Recommended reporting

Report the following with any model result:

- OligoVigil release DOI: `10.5281/zenodo.20633779`
- task name and version
- split file checksum
- evaluation split
- metric definition
- whether Grade C records were excluded
- whether source-paper by molecule/cohort grouping was preserved

## Checksums

| file | bytes | SHA256 |
|---|---:|---|
| benchmark_reference_splits.csv | 146241 | 5bf1a917a6acd47ddd0d087723737286873b33c6e7062db5a0009345535079fe |
| benchmark_task_cards.csv | 1520 | 07458fa9ab82396d694c7abebae9b62b074f87b5fb2648bec24346b808120bf0 |
| benchmark_baseline_results.csv | 5125 | 4803c43392607d15fa40e332745753b291a90a17dfa92706014c53f8933af8e1 |
| evidence_release.csv | 620435 | 1e75030b03e14f40e87a445af6c96e4b0e189abf78604998767cfe2baa78c775 |