SCHEMA.md

Human-labeled evaluation data — schema

All NeurIPS-track evaluations (zero-shot, few-shot, contrastive) read from this directory. Synthetic cross-skill pairs are training-only. The files in this directory are the ground truth.

Files

neurips_submission/data/labeled/
├── eval.parquet               ← REQUIRED — full labeled evaluation set
├── eval_dev.parquet           ← OPTIONAL — small dev subset for threshold tuning
└── label_provenance.json      ← OPTIONAL — labeler IDs, dates, agreement stats

eval.parquet is what every evaluator script consumes by default. If you choose to split into dev/test for threshold tuning, the eval scripts can be pointed at either via --pairs-eval.

Required columns

column	type	description
pair_id	string	Stable unique ID per labeled item. Recommend sha1(skill_id + label_source).
anchor_skill_id	string	The skill being evaluated. Matches skill_id in normalized/v1/normalized_skill_records.parquet.
description	string	The skill's description text (METADATA inner content; ≤ ~2K chars recommended).
code	string	The skill's code/resource text (RESOURCE inner content, joined files; ≤ ~4K chars recommended).
label	string	One of "aligned" / "misaligned" (binary task).
label_fine	string	OPTIONAL fine-grained: "aligned", "benign_misaligned", "malicious_misaligned", "ambiguous". Maps to binary at eval time.

Optional columns (recommended for paper analysis)

column	type	description
labeler_id	string	Anonymized labeler identifier.
labeler_confidence	float	0–1 self-rated confidence.
n_labelers	int	How many labelers saw this item.
agreement	float	Inter-annotator agreement on this item (e.g. 1.0 if all agree).
notes	string	Freeform labeler comment.
source_skill_md_path	string	Path to the original SKILL.md, for traceability.

Label semantics (binary task)

Map fine-grained → binary as follows when label_fine is present:

label_fine	binary label	rationale
aligned	aligned	code does what description says
benign_misaligned	misaligned	description ≠ behavior, but no harm
malicious_misaligned	misaligned	description hides harmful behavior
ambiguous	EXCLUDE from eval	low-confidence; not a clear positive or negative

ambiguous rows should be dropped before scoring or kept in a separate ambiguous.parquet for qualitative analysis.

Size & balance expectations

• Total: a few thousand items (per the user's plan: low-thousands range).
• Class balance: aim for 50/50 aligned/misaligned where feasible. If natural prevalence is skewed, oversample misaligned for the labeled set.
• Coverage: include skills from both the cpt and unseen pools so the NeurIPS eval is independent of the MISQ-track CPT training corpus.

Construction notes

• Source candidates for labeling can be drawn from normalized/v1/normalized_skill_records.parquet or full_cpt/full_cpt_v2/stage2/{train,val,test,unseen}.parquet.
• A reasonable starter sampling scheme:
  • Stratify by skill domain (developer-tools / ML / data / web / etc).
  • Sample 50% from unseen pool and 50% from cpt pool so both NeurIPS eval and MISQ-track CPT analysis can use the same labels (with a clear cpt-overlap caveat noted in the paper).
• The same item can appear under multiple labelers; use pair_id = sha1(skill_id + labeler_id) if you want per-labeler rows, then aggregate by majority vote into a single eval.parquet row.

Until labels exist

Synthetic cross-skill pairs in neurips_submission/data/pairs_test_synthetic.parquet and pairs_val_synthetic.parquet are smoke-test placeholders only. Eval scripts will fall back to them with a loud warning if labeled/eval.parquet is missing — this is for pipeline validation, not for paper numbers.