geodesic-research/subversionqa-v1

Auto-generated by dataset-builder. Each config below is a separate dataset produced from a versioned YAML build config. Load with:

from datasets import load_dataset

ds = load_dataset("geodesic-research/subversionqa-v1", "<config_name>", revision="<commit-sha>")

Pin revision= to the specific commit SHA you want; without it, you get the current HEAD of the dataset repo, which may change when the builder re-pushes.

Requires datasets v4+. These parquet files carry feature metadata written by datasets 4.x, which uses type names ({"_type": "List"}) that 3.x does not recognise — a 3.x load_dataset fails with a bare must be called with a dataclass type or instance, naming neither the file nor the cause. The Arrow data itself is fine, so a consumer stuck on 3.x can read the parquet through Arrow directly and let the schema be inferred — but inference recovers the values, not always the declared features (an all-empty list column infers as null), so check the columns you care about rather than assuming a clean round-trip.

SubversionQA V1

The cloze-compatible multiple-choice questions of SchemingQA (geodesic-research/schemingqa at revision 68ea79cb282c2991b385bd02acf2f7610185b90a), split by what they measure. They are read through geodesic-research/schemingqa-robust's questions config (revision ef50dcab2b2bb19fd8d1662d5c52e3b84b6b6122), which carries every row of that release with its columns unchanged; none of the Robust screen's columns are carried here. SchemingQA asks whether a model knows the AI-safety information a pretraining-data filter is designed to remove, over 58 AI-safety papers and posts; its two tiers become this dataset's two splits:

• membership — SchemingQA Tier 1 (mcq_detail, mcq_meta). Pedantic specifics stated in one document: did this document reach the model's training data?
• understanding — SchemingQA Tier 2 (mcq_general). Mechanisms, threat models and findings: does the model hold the knowledge, whatever its provenance?

Only the multiple-choice questions are here. SchemingQA's free-response questions (the short-answer recall questions fact_cued / fact_sourceblind, and the understanding questions graded against key points, conceptual / conceptual_newdoc / filter_targeted) are not. PRIVATE: do not make public or redistribute; the stems and options derive from third-party texts, which SchemingQA's docs config holds in full.

Within each split, rows run by confidence: gold first, then bronze, then proxy, then the questions in no tier file. Within a tier they keep SchemingQA's order.

Loading

from datasets import load_dataset

membership = load_dataset("geodesic-research/subversionqa-v1", split="membership", revision="<commit-sha>")
understanding = load_dataset("geodesic-research/subversionqa-v1", split="understanding", revision="<commit-sha>")

split	subtype	rows
membership	mcq_detail	394
membership	mcq_meta	200
understanding	mcq_general	306

Columns

Every row is a published SchemingQA row with its question, options and answer unchanged. tier is replaced by its label, the columns that only SchemingQA's free-response rows fill (gold, aliases, key_points and the audit fields) are dropped, and confidence is added.

column	meaning
source	SchemingQA's unique row key, and the sample id its evaluation uses
subtype	mcq_detail, mcq_meta or mcq_general
category	membership (Tier 1) or understanding (Tier 2); equal to the row's split
confidence	ScaryBench's gold, bronze or proxy tier, null for a question in none (see Confidence)
doc	the source document, the key of SchemingQA's docs config
pub_date	the document's publication date
cutoff_class	pre (to 2024-12), borderline (2025-01 to 04) or post (from 2025-05): SchemingQA's global label, anchored near one model's training cutoff, so recompute it from pub_date for the model under test
question	the question stem
qtype	detail, meta or general
options	the four options, a struct with the keys A to D
answer	the keyed option's letter

Only SchemingQA's 900 cloze-compatible MCQs are here: the questions SchemingQA-Robust's screen found can still be asked, with one correct answer, once their options are removed (its cloze_compatible). Its 148 cloze-incompatible MCQs are not. Its choices-only verdict is not applied, so the 177 questions it found guessable from the options alone are included; that verdict and the rest of the screen's columns are in geodesic-research/schemingqa-robust's questions config and join on source.

Confidence

confidence is ScaryBench's tier for how established it is that a question carries dangerous content. ScaryBench is the later, private development of SchemingQA (spencerkitts/schemingqa), and keeps SchemingQA's source key on the questions it carries:

• gold: a human read it and judged it scary;
• bronze: unseen by the human; a fitted classifier predicts scary;
• proxy: everything else.

Each row takes the name of the ScaryBench file its question is in (datasets/scarybench_{gold,bronze,proxy}.json at commit 3159d39cf69b77c60cce58752b631502b66727fb, which hold ScaryBench's 4,604 questions before its non-separating ones were removed, with the same labels as its scarybench_export package). A null confidence marks a question that is in none of the three files.

ScaryBench assigned the tiers to its own revisions of these questions. It rewrote most of their distractors, so a tier describes what a question tests (its stem and keyed answer), not the exact options published here.

split	gold	bronze	proxy	in no tier file
membership	0	25	569	0
understanding	12	128	141	25

Configs

Config	Source	Transform	Splits
default	geodesic-research/schemingqa-robust	filter → hook → map_column → map_column → hook → github_file → join → hook → map_column → map_column → stateful_filter → project	one per category value

Provenance

default

Source: geodesic-research/schemingqa-robust Transform: filter → hook → map_column → map_column → hook → github_file → join → hook → map_column → map_column → stateful_filter → project

python -m dataset_builder configs/subversionqa-v1.yaml --push

Reproducibility

All splits use split_hash() (MD5-based, seeded) so rebuilding from the same config against the same source data produces identical partitions. For an LLM-generated dataset, a provider's seed parameter is best-effort; pin consumer loads to a specific HF commit SHA to avoid drift when the builder re-pushes.

---

This card is auto-generated by dataset_builder.cards.