README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

finance-data — DHR / ILMN M&A Analysis Tasks

Investment-banking analysis tasks set in a (fictional) strategic M&A project where Danaher (DHR) explores acquiring Illumina (ILMN). Each task gives an agent a prompt plus a full deal-room of reference materials (financial models, spreadsheets, PDFs, research) and grades the answer against a per-criterion rubric. Tasks that require modifying a workbook also ship the expert's gold workbook as a deliverable file.

Layout

data/                 # the task table (parquet, 3 rows)
reference_folder/     # the shared deal-room every task works from
├── filesystem/       #   models, filings, research, deal docs (xlsx/pdf/pptx/…)
└── .apps_data/       #   app state (mail, calendar, chat) for the simulated environment
deliverable_files/    # gold deliverable workbooks, one subfolder per task

All tasks share one reference folder — they are set in the same deal. The reference_folder column holds its repo-relative path.

Columns

Column	Type	Description
task_id	string	e.g. task_helix_dhr_ilmn_06
sector	string	Finance and Insurance
occupation	string	Investment Banking Analyst
prompt	string	Self-contained task instructions (includes scenario framing)
reference_folder	string	Repo-relative path to the shared deal-room (reference_folder)
deliverable_files	list[string]	Repo-relative paths to the gold deliverable workbook(s); empty for answer-only tasks
gold_response	string	The expert gold answer (text)
rubric_pretty	string	Human-readable rubric
rubric_json	string	JSON-encoded rubric (see below)

Rubric format (rubric_json)

A JSON array; each entry is one grading criterion with fields:

• verifier_id, index — identity/ordering
• criteria — the criterion text
• criteria_explanation — step-by-step reference reasoning for the grader
• weight — contribution to the weighted task score
• is_primary_objective — a task passes only if all primary criteria pass
• tolerance.type / tolerance.low / tolerance.high / tolerance.units — numeric acceptance band (e.g. exact, absolute_range)
• dependencies — criteria that must pass for this one to be meaningful
• criterion_type, evaluation_scope, expected_file_type, human_rating, tags, universal

Intended harness

Populate a sandboxed environment from reference_folder, run the agent with file/spreadsheet/PDF tools, and grade the final answer with an LLM judge that applies each rubric criterion within its stated tolerance. A task passes iff every primary-objective criterion passes. For tasks with deliverable_files, the gold workbook shows the expected end-state of the modified model.