DaedalusBench-Preview / reference_definitions.md
HZPM's picture
DaedalusBench preview 1.0.0
46fbad9
|
Raw History Blame Contribute Delete
4.76 kB

Reference definitions

Each task directory tasks/<task>/ holds a Harbor task (task.toml, instruction.md, environment/, tests/), a case study (README.md) and one self-contained public contract, eval_config.json. This file defines that contract, the answer shapes and the reward.

eval_config.json

  • id: the task identifier, equal to the directory name and to the Harbor task name without the prima-mente/ prefix.
  • eval_uuid: a UUID for this released task instance.
  • version: the package version.
  • title, domain: the public title and domain shown in tasks.csv.
  • task: the prompt shown to the agent, identical to instruction.md.
  • environment: the task image (docker_image, image_file, image_id), the read-only data directory (data_dir, always /data), the answer directory (output_dir, always /workspace/output), the network policy (none), and the CPU, memory and wall-clock budget of one episode.
  • data_files: every file the agent can read under /data, with its size in bytes.
  • answer: the files the agent must write to output_dir and the graded fields. A free-text reasoning field (or a report.md) is requested by most prompts and is not graded.
  • ground_truth: the reference value of every graded field. For onek1k-age-transfer and preeclampsia-cfrna-transfer, whose fields are computed from the agent's own predictions, it points to the withheld outcomes (tests/hidden/ages.csv, tests/hidden/labels.csv) and gives the cohort summaries derived from them.
  • grader: type (always all_fields_within_tolerance), implementation (the files that grade the task), and config, one rule per graded field (see below).
  • canary: the contamination canary string.

Grading rules (grader.config)

Every rule must pass for the task to pass.

  • absolute_tolerance: the submitted value must be a finite number and satisfy |submitted - ground_truth| <= tolerance.
  • max_metric: a metric computed from the submission against the ground truth must not exceed tolerance (metric names it, for example the mean absolute error of the submitted ages against the withheld ages).
  • min_metric: a metric computed from the submission against the ground truth must be at least threshold (for example the AUROC of the submitted probabilities on the withheld outcomes).
  • interval_covers_realised: the submitted [low, high] must contain the realised value named in reference and be no wider than max_width.
  • absolute_tolerance_vs_realised: like absolute_tolerance, but the reference value is computed from the agent's own submission and the ground truth (for example, the realised mean absolute error of the submitted ages, their least-squares slope on the true ages, or the realised AUROC of the submitted probabilities).

Values in ground_truth and grader.config are stored at full precision; the case studies show them rounded.

Answer shapes

onek1k-naive-cd8, cfrna-preeclampsia-placental and snisor-coupling expect one JSON object in /workspace/output/answer.json whose keys are the graded fields plus reasoning:

{
  "<field_name>": 0.0,
  "reasoning": "<brief account of the analysis>"
}

preeclampsia-cfrna-transfer expects predictions.csv (sample_id,prob_preeclampsia, one row per withheld sample, probabilities in [0, 1], not all identical) and answer.json:

{"auroc_estimate": 0.0, "auroc_ci95": [0.0, 0.0]}

with 0 <= low <= auroc_estimate <= high <= 1. Numbers must be JSON numbers (null is not accepted); extra keys are ignored. onek1k-age-transfer expects predictions.csv (donor_id,predicted_age, one row per withheld donor), pool_means.csv (pool_id,mean_age, one row per withheld pool) and estimate.json (mean_age_estimate, sd_age_estimate, mae_estimate, slope_estimate; all but the slope non-negative); every withheld donor and pool must appear exactly once.

Reward

tests/test.sh runs tests/verify.py against /workspace/output and the files in tests/hidden/, writes the full result to /logs/verifier/result.json and the reward to /logs/verifier/reward.json:

  • reward.json holds one key, reward: 1.0 if every graded field passes its rule, else 0.0.
  • result.json holds the per-field pass flags and errors, for diagnosis only.

A missing file, unreadable JSON, a missing field or a non-numeric value gives reward 0. Grading is deterministic: no randomness, no network and no model calls. verify.py needs only Python 3.8 or newer and its standard library, so it can also be run outside the image:

python3 tasks/<task>/tests/verify.py --submission <output dir> --hidden tasks/<task>/tests/hidden --out result.json