playgent-hf/investment-banking

Playgent-IB-Bench

A benchmark dataset for evaluating AI agents on real-world investment banking tasks. Each task places an agent inside a simulated deal environment (a ZIP archive containing financial documents, Excel templates, and supporting data) and asks it to complete analyst-level work products.

Dataset Structure

test.jsonl          # 40 tasks with prompts and world assignments
verifiers/           # Per-task JSONL files with graded verification checks
worlds/              # Environment ZIP archives containing deal documents

Usage

You can use this dataset to run evals and benchmark model performance. Our RL environment is available via Prime Intellect, a hub for building and sharing RL environments.

Prime Intellect (Recommended)

We host our RL environments on Prime Intellect — the easiest way to run tasks and evaluations against this dataset.

• Environment: playgent/playgent-investment-banking
• Docs: docs.primeintellect.ai

Manual Harness

import json

# Load tasks
tasks = []
with open("train.jsonl") as f:
    for line in f:
        tasks.append(json.loads(line))

# Load verifiers for a task
task_id = tasks[0]["task_id"]
verifiers = []
with open(f"verifiers/{task_id}.jsonl") as f:
    for line in f:
        verifiers.append(json.loads(line))

Tasks

Each line is a task:

{
  "task_id": "task_xatm_ma_1_adj_ebitda",
  "world_id": "env_xatm_ma_1",
  "prompt": "..."
}

Verifiers

Each line is a graded check against the agent's output:

{
  "task_id": "...",
  "name": "check_name",
  "type": "spreadsheet_cell" | "llm_judge",
  "file": "path/to/file.xlsx",
  "sheet": "Sheet1",
  "cell": "C8",          // for spreadsheet_cell
  "expected": 12345,
  "tolerance": 0.01,
  "criteria": "...",     // for llm_judge
  "reward": 3            // positive or negative
}

Verifier types:

• spreadsheet_cell: checks a specific cell value against an expected result with a numeric tolerance
• llm_judge: uses an LLM to grade a binary YES/NO criterion (e.g. structural correctness, formula conventions)

Rewards can be negative (penalizing common pitfalls).

Worlds

Each world is a ZIP archive containing the deal environment — source documents, financial data, and Excel templates the agent must work with.

World	Description	Task Count
env_xatm_ma_1 / env_xatm_ma_1.5 / env_xatm_ma_2	ATM portfolio business — buy-side M&A	7
env_ge_spinoff	GE Aerospace spinoff — public comps analysis	2
env_juniper_sellside	Juniper Networks — sell-side process	7
env_kering_takeover_1 / env_kering_takeover_2	Kering — leveraged takeover / LBO	9
env_take_private_reit / env_take_private_reit_2	REIT take-private transaction	15

Citation

If you use this dataset, please cite:

@dataset{playgent_investment_banking_2025,
  title  = {Investment Banking Agent Benchmark},
  author = {Playgent},
  year   = {2026},
  url    = {https://huggingface.co/datasets/playgent-hf/investment-banking}
}