Loading dataset details…
harbor run -d valenlabs/frontierbiomlharbor run -d valenlabs/frontierbiomlFrontierBioML: four machine learning tasks in computational biology (binder triage, decoy scoring, partner-selective KRAS variants, budgeted screening) with deterministic held-out grading. Each task includes a reference solution; no sealed data.
harbor run -d valenlabs/frontierbiomlFrontierBioML: four machine learning tasks in computational biology (binder triage, decoy scoring, partner-selective KRAS variants, budgeted screening) with deterministic held-out grading. Each task includes a reference solution; no sealed data.
harbor run -d valenlabs/frontierbiomlFrontierBioML evaluates coding agents on machine learning tasks in computational biology. Each task gives an agent real experimental data and asks it to write a prediction script for measurements it has never seen. Grading re-runs that script on a held-out split in a sealed, network-free sandbox, so a score reflects what the script generalises to, not what the agent saw.
Results and method: https://www.frontierbioml.com
This dataset is the four preview tasks. The full set of 25 is available on request.
| Task | What the agent does |
|---|---|
frontierbioml-t01 |
Predict VHH binding to IL-6 variants and rank a family-diverse binder panel |
frontierbioml-t09 |
Score docking poses so the top pose in each panel is near-native |
frontierbioml-t22 |
Predict KRAS folding and partner binding, and pick partner-selective variants |
frontierbioml-t25 |
Spend a 700-reveal budget over three screening rounds, then commit a panel |
Each task is pinned here by content digest.
harbor dataset download valenlabs/frontierbiomlEach task ships its reference solution (solution/; for t25, one per step under steps/), which harbor run -a oracle
runs. It reads only what an agent has.
Held-out labels, sealed grading data and the scorers. Grading happens on the host.
These packages cannot score themselves. The held-out labels and the scorers stay on the host, so the answers do not ship with the questions.
A public evaluation split is in preparation: a second held-out set per task, disjoint from the one behind the reported numbers, published so you can grade your own runs offline. Until then, ask us for a grade.
Each task redistributes third-party data under its own licence; DATA-SOURCES.md carries the full
notices and the changes made. In short: t01 AVIDa-hIL6 (COGNANO, CC BY-NC 4.0, noncommercial only);
t09 DeepRank-Ab (Zenodo 17911452, CC BY 4.0); t22 KRAS deep mutational scan (Weng et al., Nature 626,
643-652, 2024, CC BY 4.0, plus six RCSB PDB entries); t25 Bits to Binders (Zenodo 20451404, CC BY 4.0).
FrontierBioML evaluates coding agents on machine learning tasks in computational biology. Each task gives an agent real experimental data and asks it to write a prediction script for measurements it has never seen. Grading re-runs that script on a held-out split in a sealed, network-free sandbox, so a score reflects what the script generalises to, not what the agent saw.
Results and method: https://www.frontierbioml.com
This dataset is the four preview tasks. The full set of 25 is available on request.
| Task | What the agent does |
|---|---|
frontierbioml-t01 |
Predict VHH binding to IL-6 variants and rank a family-diverse binder panel |
frontierbioml-t09 |
Score docking poses so the top pose in each panel is near-native |
frontierbioml-t22 |
Predict KRAS folding and partner binding, and pick partner-selective variants |
frontierbioml-t25 |
Spend a 700-reveal budget over three screening rounds, then commit a panel |
Each task is pinned here by content digest.
harbor dataset download valenlabs/frontierbiomlEach task ships its reference solution (solution/; for t25, one per step under steps/), which harbor run -a oracle
runs. It reads only what an agent has.
Held-out labels, sealed grading data and the scorers. Grading happens on the host.
These packages cannot score themselves. The held-out labels and the scorers stay on the host, so the answers do not ship with the questions.
A public evaluation split is in preparation: a second held-out set per task, disjoint from the one behind the reported numbers, published so you can grade your own runs offline. Until then, ask us for a grade.
Each task redistributes third-party data under its own licence; DATA-SOURCES.md carries the full
notices and the changes made. In short: t01 AVIDa-hIL6 (COGNANO, CC BY-NC 4.0, noncommercial only);
t09 DeepRank-Ab (Zenodo 17911452, CC BY 4.0); t22 KRAS deep mutational scan (Weng et al., Nature 626,
643-652, 2024, CC BY 4.0, plus six RCSB PDB entries); t25 Bits to Binders (Zenodo 20451404, CC BY 4.0).