File size: 4,918 Bytes
46fbad9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | ---
license: other
license_name: daedalusbench-preview-license
license_link: LICENSE
pretty_name: DaedalusBench-Preview
language:
- en
size_categories:
- n<1K
tags:
- biology
- genomics
- single-cell
- immunology
- cell-free-rna
- long-read
- benchmark
- agents
- reasoning
- scientific-inference
- rl-environment
- harbor
configs:
- config_name: tasks
data_files:
- split: release
path: tasks.csv
extra_gated_heading: Request access to the DaedalusBench preview
extra_gated_prompt: >-
DaedalusBench is shared as a preview under the DaedalusBench Preview Licence (see LICENSE): evaluation of AI
systems only, no redistribution, and no training on the tasks, answers or graders.
extra_gated_fields:
Name: text
Organisation: text
Intended use: text
I agree to the DaedalusBench Preview Licence: checkbox
extra_gated_button_content: Request access
---
# DaedalusBench-Preview
**Preview · 5 tasks · October 2026**
DaedalusBench is built on Prima Daedalus. It turns the data we use to train and evaluate our biology models into RL
environments in which LLM agents carry out analyses that our human scientists do today. Every task and every check
was annotated and reviewed by human scientists and follows a real-world scientist workflow.
Each task gives an agent a large, messy dataset built from a public biomedical deposit, in a sandbox without network
access, and asks for a few quantities that only a correct end-to-end analysis recovers. A deterministic grader gives
reward 1 only if every graded field is within its tolerance.
## Tasks
| # | Task | Title | Domain | Graded fields |
|--:|---|---|---|--:|
| 1 | [`preeclampsia-cfrna-transfer`](tasks/preeclampsia-cfrna-transfer/README.md) | Preeclampsia prediction in a new cfRNA cohort | Cell-free RNA / prenatal | 3 |
| 2 | [`onek1k-age-transfer`](tasks/onek1k-age-transfer/README.md) | Donor ages in a withheld single-cell cohort | Single-cell RNA-seq / immune ageing | 6 |
| 3 | [`onek1k-naive-cd8`](tasks/onek1k-naive-cd8/README.md) | Naive-CD8 ageing effects under contamination | Single-cell RNA-seq / immune ageing | 4 |
| 4 | [`cfrna-preeclampsia-placental`](tasks/cfrna-preeclampsia-placental/README.md) | Placental cfRNA share in preeclampsia and a screening projection | Cell-free RNA / prenatal | 7 |
| 5 | [`snisor-coupling`](tasks/snisor-coupling/README.md) | Exon coupling in single-nucleus long reads | Long-read single-nucleus RNA-seq / splicing | 6 |
Each task directory holds a case study (`README.md`), the prompt and the grader. The dataset viewer shows
`tasks.csv`: one row per task with the full prompt.
## Results
| Model | Pass rate, % | Field score, % | Wall time, min | Cost, $ | Tokens in / out |
|---|--:|--:|--:|--:|--:|
| Claude Opus 5.5 (medium) | **4.0** | 41 | 3.2 | 0.41 | 138k / 13k |
| Gemini 4 Argon (default thinking) | **4.0** | 23 | 30.2 | 1.26 | 2.8M / 64k |
| GPT-6-astra (medium) | **0.0** | 44 | 11.2 | 2.58 | 1.1M / 26k |
- **Pass rate:** share of rollouts with every graded field within tolerance, averaged over the five tasks (10 rollouts per task and model); rollouts that hit the 30-minute budget count as failures.
- **Field score:** share of graded fields within tolerance (partial credit), averaged the same way.
- **Wall time:** median minutes per rollout.
- **Cost:** median list-price cost per rollout.
- **Tokens:** median per rollout; input includes cached tokens, output includes reasoning tokens.
## Package layout
```text
<repo-root>/
├── README.md, LICENSE, tasks.csv, manifest.json, checksums.sha256, reference_definitions.md
├── tasks/<task>/ # case study, eval_config.json, instruction.md (prompt), task.toml, environment/, tests/ (grader)
└── images/<task>.tar.zst # task image with the agent-visible data under /data
```
## Running a task
```bash
hf download prima-mente/DaedalusBench-Preview --repo-type dataset --local-dir daedalusbench --exclude "images/*"
hf download prima-mente/DaedalusBench-Preview --repo-type dataset --local-dir daedalusbench --include "images/<task>.tar.zst"
cd daedalusbench && zstd -dc images/<task>.tar.zst | docker load
harbor run -p tasks/<task> -a <agent> -m "<model>" -y
```
Each episode needs 16 CPUs, 96 GB RAM and no network. `tests/` holds the ground truth, so keep it away from the agent.
`reference_definitions.md` documents `eval_config.json` and the reward.
## Licence
© 2026 Prima Mente, shared under the DaedalusBench Preview Licence (`LICENSE`): evaluation only, no redistribution, no
training.
## Citation
```bibtex
@misc{daedalusbench2026,
title = {DaedalusBench: agentic scientific-inference tasks in computational biology},
author = {{Prima Mente}},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/prima-mente/DaedalusBench-Preview}},
note = {Preview, 5 tasks}
}
```
## Contact
Prima Mente · hongjian@primamente.com
|