You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
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.
Log in or Sign Up to review the conditions and access this dataset content.
SKJudge-Promotion
What this is. An editorial-judgment environment built from the live promotion record of Seek, an autonomous research agent: every night a "promotion judge" session decides which captured claims enter her permanent vault and which are declined, with reasons. Here your agent sits in that chair.
Analogy. A newspaper's night editor's desk, with the newspaper's actual decision log as the answer key — and, separately, the fact-checker's report from three months later as a second, harsher answer key.
Two families = two golds (deliberately different questions):
agree-*(200 tasks x 2 tiers): given a capture (all verdict frontmatter stripped) and a shuffled candidate list, rule promote/decline per candidate. Gold = what the historical judge actually ruled. Reward = balanced accuracy (all-promote scores 0.5). This measures AGREEMENT with a particular judge's taste — including her mistakes. It is a behavioral-cloning target, and the card says so out loud.outcome-*(100 tasks x 2 tiers): given a capture and its claims with quotes and sources, bet promote_now vs hold_for_verification. Gold = what mechanical verification later PROVED: all quotes verbatim => promote_now; claims unsupported by the very source she archived => hold. Quote-broke-but-claim-supported cases (extraction artifacts) count as promote_now — transcription debt is not bad judgment. This measures judgment against reality. The two scores can disagree; that disagreement is the most interesting number this environment produces.
Two tiers. -t1 = single-shot (everything in the instruction).
-t2 = the same task with the capture's 1-hop vault neighborhood mounted
at /app/vault for the agent to read. The t2-minus-t1 delta measures what
context-gathering is worth to judgment.
Baselines. agree: 0.5 by construction for any constant strategy. outcome: the gold split is 100 tasks at roughly 53/47, so constant answers sit near 0.5. Beat both before claiming editorial judgment.
The full record. The runnable agree tasks are a deterministic
200-capture sample; agreement-full.jsonl carries ALL qualifying captures
(candidates + gold) for analysis. tasks_index.jsonl is the browsable
index.
Limits, honestly. Agreement gold is one agent's taste, curated by one human's specs — high agreement means you judge LIKE SEEK, not that you judge well; her outcome-refuted keeps are in there as correct answers. Outcome gold is asymmetric: only promoted captures have downstream verification, so declines are never outcome-scored, and the artifact-vs-unsupported split leans on a 7B grounding model's triage (chunk-local; absence-claims read as unsupported by construction). Verdict-bearing frontmatter is stripped from every agent-visible file and mechanically leak-checked to zero; 296 decline candidates whose one-line stubs still implied their verdicts were dropped rather than laundered.
Terms. Research and evaluation ONLY. Captures and verdicts were written by Claude-family models: do not use this environment or its contents to train models. Same gate as every seekbot dataset.
Generated mechanically by seek_export_judgment_env.py (seek-code);
no model wrote any part of this export.
- Downloads last month
- 11