직업을 없애는 질문 대신 업무를 다시 설계하기

from datasets import load_dataset


worker_desire = load_dataset("SALT-NLP/WORKBank", data_files="worker_data/domain_worker_desires.csv")["train"]

expert_ratings = load_dataset("SALT-NLP/WORKBank", data_files="expert_ratings/expert_rated_technological_capability.csv")["train"]

task_meta_data = load_dataset("SALT-NLP/WORKBank", data_files="task_data/task_statement_with_metadata.csv")["train"]

task_summary = (
    domain_worker_desires
    .groupby("Task")
    .agg({
        "Automation Desire Rating": "mean",
        "Job Security Rating":    "mean",
        "Enjoyment Rating":       "mean",
    })
    .reset_index()
)

desire_by_task = domain_worker_desires.groupby("Task")["Automation Desire Rating"].mean().rename("automation_desire")
capacity_by_task = expert_ratings.groupby("Task")["Automation Capacity Rating"].mean().rename("automation_capacity")

task_scores = pd.concat([desire_by_task, capacity_by_task], axis=1).dropna().reset_index()
# Compute correlation between automation capacity and desire
from scipy.stats import spearmanr

r_s, p_s = spearmanr(task_scores["automation_capacity"], task_scores["automation_desire"])

- Each task lives at `dataset/<split>/<profession>/taskN/` and contains:
  - `task_folder/` - the working directory the agent operates in
  - `RUBRICS.json` - weighted pass/fail criteria used for scoring
  - `task_card.md` - human-readable task brief (sourced from ONET; not seen by the agent)
- **Supported agents**: Claude Code, Codex CLI, OpenCode
- **Judge**: any OpenAI-compatible endpoint; default is `grok-4.3` on xAI.