Epsilab/accounting-tasks-sample

Epsilab Accounting Tasks (Sample)

This is a sample of five expert-written accounting tasks from Epsilab. Each task is in the Harbor task format. Each task comes with its grader and the solver trajectories from its most recent run.

An accountant wrote each task against the books of a small business. The agent works in a Linux desktop with Google Drive, Google Sheets and Box. It must deliver a work product, such as a reconciliation workbook or a review memo. A model judge then grades the deliverable against the expert's binary criteria.

The source documents (bank statements, ledgers, reports) are not included.

Summary

Item	Count
Tasks	5
Attempts (trajectories)	20 (4 for each task)
Criterion grades	256
Solver	gpt-6-astra, high reasoning effort, up to 300 steps
Judge	gpt-6.1-sol, high reasoning effort, 3 samples for each criterion

Task	Criteria	Mean reward
Cohen Rush Pugh 2026 Cash Review	12	0.308
Cohen Rush Pugh 2024 Year End Review	15	0.288
Cochran & Stephens – Q1 2026 Bank Reconciliation Exceptions and Correcting Entries	15	0.311
Fixed Assets and Depreciation Review 2024–2026	12	0.433
Cameron Madison of Anne LLC: Lender Debt Reconciliation and Coverage Ratios	10	0.527

Files

tasks/<task>/                      Harbor task, as packaged for the run
  task.toml                        task, agent, verifier and environment settings
  instruction.md                   the instruction the agent receives
  tests/judge_prompt.md            the grader prompt
  tests/reward_criteria.toml       the criteria, weights and judge settings
  tests/test.sh                    verifier entry point (runs Harbor Reward Kit)
trajectories/<task>/<trial>/
  trajectory.json                  full agent trajectory (ATIF v1.7)
  images/                          screenshots that trajectory.json refers to
  outputs/                         the deliverables the agent saved
  result.json                      Harbor trial result
  verifier/reward.json             final reward
  verifier/reward-details.json     per-criterion verdicts with the judge's reasoning
  verifier/judge-samples.json      the three judge samples
data/
  tasks.jsonl                      one row for each task
  trajectories.jsonl               one row for each attempt
  criterion_grades.jsonl           one row for each criterion of each attempt

The environment/ folder of each Harbor task is not included. It refers to private container images and to the seed data.

Grading

Each criterion is a binary check (pass or fail) with a weight of 1, 3 or 5. The judge reads the agent's deliverables and trajectory. It grades each criterion three times. The majority verdict counts. The reward is the weighted mean of the verdicts, from 0 to 1.

Tables

All three tables join on harbor_task, the folder name under tasks/.

tasks

Column	Description
harbor_task	folder name under tasks/
title	task title
instruction	the instruction text
judge_prompt	the grader prompt
criteria	list of name, description, weight, optional
n_criteria	number of criteria
solver_model, judge_model	models used
n_attempts, mean_reward	run size and mean reward
path	the task folder

trajectories

Column	Description
attempt_id, trial_name	attempt IDs; trial_name is the folder under trajectories/<task>/
harbor_task	the task
model	solver model
reward	final reward, 0 to 1
steps, tool_calls, hit_step_cap	trajectory size
started_at, finished_at	timestamps (UTC)
outputs	file names of the deliverables
final_message	the agent's last message
trajectory_path	path to the full ATIF trajectory

criterion_grades

Column	Description
attempt_id, harbor_task	the attempt and its task
criterion, weight, optional	the criterion
value	1.0 for pass, 0.0 for fail
verdict	the judge's majority answer (yes or no)
samples, agreement	the three sample answers and how much they agree
reasoning	the judge's reasoning

Loading

from datasets import load_dataset

trajectories = load_dataset("Epsilab/accounting-tasks-sample", "trajectories", split="train")
tasks = load_dataset("Epsilab/accounting-tasks-sample", "tasks", split="train")
grades = load_dataset("Epsilab/accounting-tasks-sample", "criterion_grades", split="train")

To get the Harbor tasks and full trajectories:

hf download Epsilab/accounting-tasks-sample --repo-type dataset --local-dir epsilab-sample

Notes

• Each task folder is the exact package that the run used, so the criteria match the grades.
• Contact Epsilab for access to the full dataset and the environments.