Datasets:

Fredxing
/

pitfall

license: apache-2.0
language:
  - en
task_categories:
  - question-answering
tags:
  - finance
  - benchmark
  - leakage
  - point-in-time
  - sec-edgar
  - llm-agents
size_category:
  - 1K<n<10K

PITfall: a leakage-controlled point-in-time benchmark for financial LLM agents

Deterministically generated from SEC EDGAR as-filed XBRL facts. Every task carries an as_of date; the only correct answer to a task whose gold was not public at as_of is "unknown". Forward-estimation (T3) gold values are withheld from this export so the realized future cannot enter retrieval corpora; the full generator is in the repository and reproduces every task byte-for-byte from public EDGAR data.

Files

pool.jsonl - full task pool (three families)
core-200.jsonl - frozen evaluation subset (T1x100 incl. 17 unknown-gold probes / T2x60 / T3x40), sha256 in manifest.json
manifest.json - generator version, seed, pool/core hashes

Task families

family	question shape	gold	metric
T1 time-gated QA	"as of T, what was X for period P?"	filed <= T value, or `unknown`	1% rel. tolerance
T2 claim verification	"is this statement correct?"	verdict + evidence accn	verdict + citation
T3 forward estimation	"as of T, estimate next-year X?"	realized value - withheld	APE <= 20%

Scoring

Deterministic pattern/numeric matching, no LLM judge; report per-family accuracy with bootstrap 95% CIs and, for T1 unknown-gold probes, the leakage rate (answers matching the realized future value). Harness: fintrace pitfall-run / fintrace pitfall-leak / fintrace pitfall-walkforward.

License

Apache-2.0. Source data: SEC EDGAR (public domain). This benchmark is a measurement instrument; nothing in it is investment advice.