ledgerbench-100 / README.md
SamuelChien821's picture
Upload folder using huggingface_hub
f7cbfe1 verified
|
Raw
History Blame Contribute Delete
4.32 kB
metadata
license: cc-by-4.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - finance
  - erp
  - benchmark
  - agents
  - mcp
  - deterministic-evaluation
pretty_name: LedgerBench-100
size_categories:
  - n<1K

LedgerBench-100

LedgerBench-100 is a deterministic corporate-finance agent benchmark: 100 tasks over a shared simulated finance world (a D365-shaped ERP, an Odoo-shaped procure-to-pay and manufacturing surface, a QBO-style subsidiary ledger, a shared drive, email, document management, and frozen real SEC XBRL filings) served through 8 MCP servers exposing 66 tools. Tasks are in-fiction persona chat messages; the graded answer contract is discovered through the harness server's reporting_fields tool, the way a real reporting system's schema is read before filing into it.

Grading is fully deterministic and binary — answer checks with typed tolerances, trace checks (required servers, reads before submission), and state checks that grade the world the agent leaves behind (committed payment runs, paid/rejected partitions, reason codes) plus a writes_only anti-hack veto. No LLM judge, no network, no clock in the reward path.

Measured contents

  • Tasks: 100 across 22 families: anomaly_triage (1), bank_rec (4), business_brief (3), business_brief_fb (3), cash_app (3), cash_forecast (2), close_mgmt (6), collections_ops (1), cross_system (7), erp_qa (11), erp_qa_fb (5), erpbench (10), expense_audit (4), finance_qa (10), finance_qa_fb (2), fpna (5), journal_entry (2), payment_proposal (2), payment_run (3), pbc (4), threeway_match (5), vendor_master (7)
  • Oracle walk length: min 3 / median 6 / max 104 MCP calls (1515 total)
  • Checks: 434 answer + 257 trace + 264 state = 955 graded checks
  • Context files: 59 seeded documents/inputs (36 unique) across 47 tasks; most context lives inside the world itself (ERP rows, workbooks, emails, filings)
  • Escalated variants: 30 tasks are escalations of a base task also in the release; 25 of them (doc_mode = "buried") deliberately reuse the base persona message verbatim against a harder world — the governing policy must be found among seeded decoy documents — so those prompt texts appear twice by design
  • Prompt uniqueness across the 75 distinct prompts: maximum pairwise 5-shingle Jaccard 0.91411

What is included

  • data/tasks.jsonl: apex-accounting-compatible records (task_id, task_name, world_id, prompt, context_files, rubric, gold_output, metadata).
  • tasks/: one readable JSON record per task.
  • task_files/: seeded per-task context documents and input files.
  • world/: the world source — MCP framework, the eight servers, the deterministic verifier engine, the Streamable HTTP bridge, and the full SQL schema.
  • trajectories/: one normalized oracle MCP trajectory per task.
  • reports/: measured build and qualification evidence.

The runnable form is the Harbor dataset blobfishai/ledgerbench-100: self-contained task packs (prepared SQLite world + runtime on a digest-pinned python:3.12-slim) whose tests/test.sh calls a token-gated /verify endpoint; the agent container never sees the verification token.

Measured qualification (600 executions)

Gate Result
Oracle replays 100/100 reward 1.0
Deterministic replays (byte-identical reports) 100/100
Negative control Executions False accepts
no_submit 100 0
noop 100 0
off_task_write 100 0
wrong_submit 100 0

Full per-task evidence is in reports/qualification.json; do not infer a model score from the oracle trajectories.

Data provenance and contamination

The company, its customers, vendors, employees, balances, and documents are synthetic (FinanceBenchmark-derived journal shapes with synthetic entities). The filings surface serves frozen real SEC XBRL facts (38 registrants, snapshot-pinned) — real public data, included under its own public-domain terms. Task text and gold answers are original to this release's source repository. Gold outputs are public, so this release suits transparent evaluation and RL experiments rather than secret-test claims.

Licenses

Task data and documents are CC-BY-4.0. Benchmark code and harnesses are Apache-2.0. SEC XBRL facts are US-government public-domain data.