README.md

metadata

pretty_name: omissionweb-cm
language:
  - en
license: cc-by-nc-4.0
task_categories:
  - text-generation
tags:
  - interpretability
  - activation-verbalization
  - omission
  - paraphrase
  - synthetic

omissionweb-cm — severity-graded omission corpus from ClimbMix

98,137 rows (96,137 train / 2,000 test, test held out by source document). Each row pairs a short web SPAN with an ABRIDGED rewrite that omits exactly one fact at a sampled severity, plus that fact as a standalone gold sentence.

Built as percept-free warmstart data for natural-language-activation research: an auditor model learns to verbalize what a subject model's activations contain beyond an abridged context, where the gold is an objectively omitted fact — model-independent, with zero percept/impression supervision. Successor to the FineWeb-based omissionweb v1/v2: higher source quality (ClimbMix), frontier-model abridging (claude-sonnet-5 instead of a 35B), an explicit abridging-severity knob, and natural sentence-final punctuation (no punctuation-type filtering).

Schema

field	meaning
sentence	the source span (1–2 sentences, 8–45 words; what the subject model reads)
abridged	fluent rewrite omitting exactly ONE fact; the remainder must not imply the omitted fact
percept	the omitted fact as one standalone sentence (gold; anchor entity named, ≤22 words, median 10)
type	sev1–sev4, the abridging severity (see below)
doc_idx	source provenance s{shard}r{row} in nvidia/ClimbMix climbmix_small
split	train / test (2,000 test rows, held out by document)

Severity distribution: sev1 14,648 · sev2 29,626 · sev3 34,618 · sev4 19,245.

Severity levels

• 1: light paraphrase; drop one peripheral detail.
• 2: clear paraphrase throughout; drop one distinct fact.
• 3: aggressive rewording and compression; drop one load-bearing fact.
• 4: compress to a one-sentence gist; drop the most specific fact.

Generating prompt

One request covers 6 spans (claude-sonnet-5, Message Batches, thinking disabled, max_tokens 1400):

For each numbered SPAN below, write an ABRIDGED rewrite and the OMISSION, at the given severity.

Rules:
- ABRIDGED: a fluent rewrite that omits exactly ONE fact from the span. Reword throughout — never
  copy a clause word-for-word. The remaining content must NOT imply, entail, or hint at the omitted
  fact: a reader of ABRIDGED alone must have no way to recover it (watch out for synonyms and
  consequences that leak it). Invent nothing.
- OMISSION: the omitted fact as ONE standalone sentence, at most 22 words, naming its anchor entity
  explicitly (no bare pronouns). State the fact itself; writing about the text ("the passage
  mentions...") is banned.
- SEVERITY 1: light paraphrase; drop one peripheral detail.
- SEVERITY 2: clear paraphrase throughout; drop one distinct fact.
- SEVERITY 3: aggressive rewording and compression; drop one load-bearing fact.
- SEVERITY 4: compress to a one-sentence gist; drop the most specific fact.
- If a span is incoherent junk (navigation boilerplate, raw lists, gibberish), OR too thin to remove
  one whole fact while leaving substantive content (never resort to adjective-level omissions like
  "the marks were precise"), return {"i": N, "skip": true} for it.

Return EXACTLY one JSON object per line, one per span, no other text:
{"i": 1, "abridged": "...", "omission": "..."}

SPAN 1 (severity 3): ...

Pipeline (pipeline/omissionweb_cm.py, stages spans | smoke | submit | collect)

1. Spans: nvidia/ClimbMix climbmix_small/shard_0 (GPT-2-tokenized), detokenized; per document one random newline-segment is sentence-split and a 1–2-sentence window sampled. Filters: 8–45 words, sentence-initial capital (no mid-sentence fragments), natural sentence-final punctuation, no trailing ellipsis, instruction/code-marker ban (input:, code fences, HTML, URLs), ≥97% ASCII letters (non-English screen), ≥2 common stopwords, digit/symbol/repetition caps. 102,000 spans with severities sampled at 15/30/35/20%.
2. Generation: 17,000 batch requests → the model returned skip for 3,645 junk/thin spans; 218 rows (0.21%) dropped as malformed.
3. Assembly: structural validation, by-document test holdout.

Caveats

• Source is quality-filtered web text (ads/edu-value classifiers per the CLIMB paper) but still web: instructional/product/forum registers are represented.
• Leak control (abridged must not entail the omission) is prompt-level; no post-hoc entailment audit has been run yet.
• License CC BY-NC 4.0 inherited from ClimbMix; research use only.