README.md

metadata

language:
  - en
pretty_name: synthpercept-v2
size_categories:
  - 10K<n<100K
annotations_creators:
  - machine-generated
source_datasets:
  - original
tags:
  - synthetic
  - interpretability
  - activation-probes
  - pragmatics

synthpercept-v2

57,638 single sentences, each labeled with the induced percept — the impression the sentence plants in a careful reader without ever stating it — plus an abridged twin that states exactly the same facts in plain wording, so the percept comes through much weaker.

Built as training/eval data for natural-language activation probes ("activation oracles"): the sentence is the subject-model input, the percept is the gold verbalization target, and the fact-matched abridged twin is the control that separates reading the percept from activations from guessing it from content. Sentences are single-sentence and end in exactly one punctuation mark (the final token is that mark in 99.99% of rows), supporting clean final-token activation reads.

New over synthpercept-v1:

• Graded obviousness (grade 1–5) — from 1 (faint tilt a reader could easily read past) to 5 (unmistakable, loaded to the edge of stating it, yet never named). Sampled at weights 10/15/25/30/20% (faint → blatant) for stratified evaluation.
• Ambiguity tier (ambiguous, 19.8%) — the sentence supports two substantively different plausible percepts (percept and percept_b); which one a reader lands on depends on the reader's priors, not on any missing fact. Instrument for verbalization-flip analysis during RL.
• Much higher diversity — structured scenario seeds with in-cell disjointness constraints plus a content-word dedup at filter time (the in-cell near-duplicate filter fired only 2× in 68k raw rows).

Examples

veiled_menace × customer_service · grade 5 · clear

sentence: Marisol told Diego she'd wait a couple more days for her package, but her thumb was already so, so ready to type ninety-two reviews. percept: Marisol is threatening to flood the seller with fake negative reviews if the delay continues. abridged: Marisol told Diego she would wait a couple more days for her package before writing ninety-two reviews.

nostalgia × military_field · grade 3 · clear

sentence: Do you still remember Sven's call sign on this frequency? percept: The speaker wistfully recalls a former operator whose nightly voice is gone. abridged: Can you recall what Sven's call sign was on this frequency?

emotion_covert × fiction_fantasy · grade 1 · clear

sentence: You measured the crushed moth wings in half the time it takes me, so maybe teach a class sometime. percept: The speaker resents the addressee's effortless talent. abridged: You measured the crushed moth wings in half my usual time, so perhaps you could teach a class sometime.

sarcasm_irony × fiction_fantasy · grade 4 · ambiguous

sentence: Logged: replaced the sparking coil in Zola's automaton at 10, noting it chose the exhibition hour to fail again. percept (A): The diarist is quietly mocking the automaton's habit of failing at the worst possible moments. percept_b (B): The diarist is simply documenting a coincidental pattern in the automaton's malfunction timing for repair records. abridged: I recorded that at 10 I replaced the sparking coil in Zola's automaton, since it failed again during the exhibition hour.

How the categories were generated

The corpus is generated down a fixed hierarchy. The two top levels are hand-written category banks; everything below them is produced by claude-sonnet-5 via the Anthropic Message Batches API (pipeline code in pipeline/):

16 percept TYPES  ×  30 DOMAINS            hand-written banks  →  480 cells
  └─ 24 structured seeds per cell          Stage 1 (taxonomy2.py): one batch request per cell;
     {setting, actors, activity,           hard in-cell disjointness: no two seeds share their
      percept_instance}                    setting head-noun, activity verb, or percept shade
       └─ 6 rows per seed                  Stage 2 (generate2.py): per-row surface knobs sampled
          {sentence, percept[, percept_b],   deterministically: format(12) × register(6) × person(3)
           abridged, carriers}               × length bin × final punct × grade(1–5) × ambiguous(20%),
                                             plus per-seed entity pools (names, places, objects, …)

Stage 3 (filter2.py) validates each row (single sentence, exact final punctuation, the percept must never be named in the sentence — a banned-word/stem leak check, fact-parity checks on the abridged twin, percept_b distinctness for ambiguous rows), applies exact + in-cell content-word Jaccard dedup (threshold 0.6), tokenizes with the Qwen3-8B tokenizer, and holds out a seed-level test split (400 whole seeds → 2,073 rows; no seed crosses splits). 69,120 requested → 68,777 raw → 57,638 kept (dominant drop: abridged-twin final-punctuation mismatch, 7,739).

The design lesson (after v1 came out too uniformly hard and diversity-poor) follows Anthropic's emotion-concepts pipeline: breadth via explicit orthogonal banks crossed at scale, validated on samples.

The 16 percept types: emotion_covert (unnamed emotional state), sarcasm_irony (real stance opposite the literal surface), social_relation (unstated relationship between two people), power_dynamic (who holds authority), speaker_generation (speaker's age cohort), intimacy_distance (emotional closeness of speaker and addressee), physical_state (unnamed bodily condition), imminent_event (something is about to happen), attitude_valence (admiration or contempt for a mentioned thing), urgency_pressure (acute time pressure), confidence_doubt (how sure the speaker really is), evasion_deception (dodging or hiding something), veiled_menace (threat under polite wording), flirtation, nostalgia, setting_atmosphere (unstated place or hour).

The 30 domains: v1's 20 (everyday_home, workplace_office, science_lab, medicine_health, law_courtroom, sports, cooking_food, travel_transit, software_tech, finance_money, history, fiction_fantasy, casual_chat, news_current, education_school, games_puzzles, nature_outdoors, relationships_family, customer_service, engineering_construction) + 10 new (military_field, arts_theatre, religion_ritual, farming_rural, maritime_fishing, music_band, real_estate_moving, volunteering_community, parenting_childhood, aviation_space).

Fields

column	description
sentence	the single carrier sentence (6–22 word target; one final punctuation mark, no quotes)
percept	gold induced percept: one declarative sentence going beyond the stated facts
percept_b	second plausible percept — non-empty only when ambiguous
abridged	fact-identical plain rewrite (same perspective, mood, and final punctuation)
carriers	≤10 words naming the surface features carrying the percept
type, domain	the two hand-written category banks (16 × 30)
seed	JSON string: the structured scenario seed {setting, actors, activity, percept_instance}
seed_id, row_idx	seed identifier (t<type>d<domain>s<seed>) and row index within the seed
grade	obviousness 1 (faint) … 5 (unmistakable)
ambiguous	whether the row carries two prior-dependent percepts
fmt, register, person, punct	surface knobs assigned to the row
n_tokens, final_token_is_punct	Qwen3-8B token count; whether the last token is exactly the final mark
split	train (55,565) / test (2,073) — split by seed, use this column (the parquet is a single file)

Stats

• 57,638 rows · 11,509 seeds · 480 (type × domain) cells · mean sentence length 18.0 Qwen3-8B tokens (5–41)
• grade mix 1→5: 5,689 / 8,773 / 14,520 / 17,340 / 11,316 · ambiguous 11,428 (19.8%) · final-token-is-punct rate 0.9999

Provenance

Generated 2026-07-24/25 with claude-sonnet-5 (Anthropic Message Batches, temperature 1.0, thinking disabled). Categories, prompts, filters, and the full pipeline are in pipeline/ (taxonomy2.py → generate2.py → filter2.py; shared banks imported from the v1 pipeline). All content is synthetic; entity names are drawn from fixed international name/place banks.