README.md

metadata

license: cc-by-4.0
extra_gated_prompt: >-
  This dataset is published for research and evaluation purposes only. Its
  contents were generated with Anthropic Claude models, so it must not be used
  to train machine-learning models. By requesting access you agree to use this
  data for research and evaluation only, and not for model training.
extra_gated_button_content: I agree — research/eval use only, no training
configs:
  - config_name: task-index
    data_files: tasks_index.jsonl
tags:
  - environment
  - rl-environment
  - harbor
  - agents
  - citation-verification
  - grounding
  - research

SKBench-Env

What this is. SKBench repackaged as a runnable agent environment (Harbor task format). SKBench is a naturalistic quote-grounding benchmark from Seek, a live autonomous research agent: 1405 claims, each attributing a quote to a public source URL, each mechanically labeled verbatim or no_match by fetching the source and testing normalized-substring containment. Here, instead of reading the labels, your agent must EARN them: it is dropped into a container with the claims (labels stripped), fetches every source itself, and is graded against the shipped gold labels.

Analogy. A fact-checker's inbox: a stack of quotes someone claims to have found in real documents. Some are genuine. Some never existed. Some are real but mangled in transcription (PDF extraction ate a space). Your job is to open every document and say which is which.

Worked example. Task skbench-000 includes claim-thuras-1930-patented-bass-reflex-loudspeaker-enclosure: the claim says "A.L. Thuras, a Bell Labs acoustics engineer, patented the bass-reflex (vented) loudspeaker enclosure in 1930 — a cabinet that works as a Hel..." and attributes a quote to http://acousticsresearchcentre.no/a-brief-history-of-electroacoustics-pt-7-loudspeaker-enclosures-en/. The agent must fetch that source (a PDF — pdftotext is installed), normalize both sides (NFKC, curly→straight quotes, dashes→hyphens, whitespace collapsed, lowercased), and decide whether the quote survives as a substring.

Structure. 147 tasks x ~10 claims (source-groups never split across tasks). Per task: instruction.md, environment/Dockerfile (+ claims.json, the agent's only input), tests/ (gold labels + grader, verifier-side only), solution/solve.sh (oracle, validates plumbing). Reward is fractional accuracy written to /logs/verifier/reward.txt.

Baselines. Answering no_match for everything scores ~0.6 (the corpus is 846/1405 no_match) — beat that before claiming your agent can fact-check. no_match does NOT mean false: it mixes fabrication, PDF-extraction artifacts, page drift, and paywall shells; separating those is the actual skill.

What this catches (a true story). While building this corpus, the agent noticed that reference [1] of Kaelbling, Littman & Cassandra's "Planning and Acting in Partially Observable Stochastic Domains" (Artificial Intelligence 101, 1998) — the paper that carried POMDPs into mainstream AI planning — dates Karl Åström's foundational paper as 1995. It was published in 1965, thirty years earlier. The agent then traced the error's provenance: the authors' own 1994 AAAI paper and their February 1998 preprint date it correctly; the misdate exists only in the final typeset journal version and every mirror of it since — a production-stage error, faithfully propagated for decades. Four machine-verified receipts for this sit in the parent dataset (claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin and its siblings). Nobody asked for that check. It fell out of an agent that reads references and verifies quotes — which is exactly the capability this environment measures.

Limits & time-sensitivity. Sources are live URLs, so this environment decays: labels were made on the per-claim verify_date, and a page edited or removed since can strand a verbatim label (instructions tell the agent to judge as-of that date). Decay is slower than you might fear — much of the corpus cites stable primaries (journal PDFs, institutional archives, arXiv, Wikipedia-class references), and in spot checks ~5 weeks after labeling, sampled tasks still replayed at 100% — but it is real: treat scores as comparable within a time window, not across years. Records carrying an evidence.sha256 in the parent dataset let future users prove whether a source has drifted since capture. PDF records were extracted with pdftotext; other extractors produce different whitespace. Claims were written by Claude-family models in Jul–Sep 2026.

Changelog. v0.1 (2026-10-02): 19 gold labels corrected from no_match to verbatim, following skbench v0.1. The original checker never ran pdftotext on extensionless arXiv /pdf/ links, so those quotes were tested against raw PDF bytes. They are verbatim under the benchmark's own rule, and their verify_date is now the re-check date, 2026-10-01. Five tasks change. Tag v0 keeps the original release.

Terms. Research and evaluation ONLY. The claim texts are Claude-model outputs: do not use this environment or its contents to train models. Same gate terms as every seekbot dataset.

Generated mechanically by seek_export_skbench_env.py (seek-code repo); no model wrote any part of this export.