Dwootton/persona-representation-corpus

Persona Representation Corpus

A reproducible corpus of persona representations — how LLM benchmarks, role-play datasets, dialogue/personalization corpora, and synthetic-user sources actually encode "a persona" — extracted from 19 sources spanning 6 representation families.

Built for the survey "What is a persona representation?"

Files

File	Contents
personas.jsonl	39,459 normalized records: source, source_repo, family, domain, generation, paper, persona_id, text, fields
field_stats.json	Field-name frequency, inferred value type (categorical/numeric/long_text/list/short_text), examples
source_summary.json	Per-source record counts, field-count and text-length stats; failures (empty — 19/19 loaded)
extractions.jsonl	3,059 personas (200/source sample) mapped by Qwen/Qwen3-4B-Instruct-2507 onto a unified 20-attribute schema, with evidence quotes, confidence, and parse status
extraction_stats.json	Per-source field-presence rates over the unified schema (layer 2)
embedding_stats.json	Within-source and cross-source mean cosine similarity, family silhouette
embeddings.npz	Per-source matrices of sentence-transformers/all-MiniLM-L6-v2 embeddings (≤400 texts/source, normalized)
extract_personas.py, llm_extract.py	Extraction scripts (layers 1 and 2) for full reproducibility
extractions.jsonl	Layer 2: unified-schema LLM extraction per persona (2,705 parsed records): 20 attributes + persona_type, strongest_evidence, confidence, plus parse_ok flag
extraction_stats.json	Per-source field-presence rates from the unified schema, model and parse-failure count
probe_results.json	Decodability: logistic probes decoding categorical attributes from persona embeddings (test acc vs majority baseline)
geometry_audit.json	Per-source embedding geometry: effective rank, top-10 variance concentration, mean pairwise cosine, centroid similarities
capacity_curve.json	Persona-as-composition capacity: held-out item retrieval (MRR/R@1) vs history length k × pooling method (mean/recency/learned attention) on 400 LaMP users
capacity_curve_v2.json	Capacity v2: three encoders (MiniLM-384, bge-base-768, bge-large-1024) × {mean, recency, attention, multi-vector M∈{2,4,8} with MaxSim}, 300-step attention, negation study per encoder

The 6 representation families

1. one_liner_role — a single sentence naming a role (PersonaHub, OpenCharacter seeds, PersonaGym). ~80–120 chars.
2. structured_profile — flat attribute records (Nemotron-Personas-USA: 22 fields; Danish: 44; Japan: 88; MatrAIx 1M: 1,290 categorical codes). Census-sampled or PGM-sampled.
3. labeled_text_profile — free text with labeled headers or atomic sentences (PersonaChat sentence lists, Synthetic-Persona-Chat prose, OpenCharacter "Name:/Age:/..." profiles, RoleBench character descriptions).
4. character_card — named-field JSON cards (Chub/SillyTavern chara_card_v2: description/personality/scenario/first_mes/mes_example; legacy TavernAI v1 naming).
5. behavioral_history — the persona IS the history (LaMP user item histories, PersonalWAB profiles from real Amazon behavior, PRISM real-human self-descriptions).
6. agent_state / persona_as_questionnaire — persona as runtime state or trait file (Generative Agents scratch.json: 48 fields incl. memory hyperparameters and daily schedules; Anthropic model-written-evals: one trait per file).

Sources

nemotron-personas-usa (2502.13123), danish-personas, japan-synthetic-personas, matrix-persona-1m, personahub (2406.20094), opencharacter-seed/card (2501.15427), personagym (2407.18416), chub-ai-cards, chub-popular-legacy, charactereval2 (2412.11912), personachat (1801.07243), synthetic-persona-chat (2312.10007), rolebench (2310.00746), prism-survey (2404.16019), lamp (2304.11406), personalwab (2410.17236), generative-agents (2304.03442), model-written-evals-persona (2212.09251).

See source_summary.json for per-source repos, domains, generation methods, and record counts.

Headline findings

• Within-source embedding similarity spans 0.135 → 0.856: one-liner LLM personas (PersonaHub, OpenCharacter seeds) are the most diverse; census-sampled structured profiles (Danish 0.856, Japan 0.727) are the most homogeneous.
• Family silhouette = 0.08: representation families are about format, not content — persona content overlaps heavily across families in embedding space.
• The demographic pentad (age, sex/gender, occupation, location, education) is the only near-universal field set, appearing as typed columns, JSON keys, and labeled headers depending on family. Personality is the most type-unstable field: Big-Five scores, free text, or a card slot.
• Text length tracks task: data-generation personas ~92 chars; dialogue personas ~135; role-play cards 400–1,900; structured/behavioral profiles ≥2,000 (truncated at 2,000 in this corpus).
• Provenance leakage is measurable: OpenCharacter seeds are statistically identical to PersonaHub personas (same mean length 91.87 chars, same distribution) — persona datasets inherit each other's generators.
• Cross-family attribute comparison (layer 2, Qwen3-4B extraction, 3,059 personas): demographics (age/gender/occupation/location) are near-universal in census-sampled and eval sources (presence ≈1.0) but nearly absent in dialogue personas (PersonaChat: age 0.055, gender 0.055) and data-generation seeds (PersonaHub: age 0.01). Psychographics dominate role-play cards (chub: personality 0.94, values 0.88) and structured profiles (1.0). Communication style (tone, linguistic habits) is the least-encoded attribute almost everywhere (PersonaChat habits 0.115; PersonaHub 0.02) — even though PersonaGym explicitly evaluates linguistic habits. History is encoded only by behavioral sources (LaMP 1.0, PRISM 0.92). Parse failure rate 11.6% (354/3,059), logged per-record.
• Task families have inverse attribute signatures (unified-schema presence rates): census/structured sources saturate demographics (age/gender/education/marital ≈ 1.0) but almost never carry scene or linguistic-habit fields (≤0.1); character cards invert this (narrative_context ≈ 1.0, linguistic_habits 0.66–0.70, demographics 0.3–0.7); dialogue personas (PersonaChat, Synthetic-Persona-Chat) carry almost no demographics (age ≤ 0.07) but high interests/values; behavioral-history personas (LaMP) carry no attribute content at all (every attribute ≤ 0.06 — the persona is the history).
• Seed→expansion adds structure, not content: OpenCharacter one-liners → labeled-header cards raise age 0.01→1.0 and gender 0.02→1.0 while leaving occupation/goals/interests rates nearly unchanged.
• Persona embeddings are partially decodable and partially invertible (probes): 6/24 source×attribute probes beat majority baseline, best matrix-persona-1m/occupation +0.217 — the interpretability/privacy-relevant leakage rate, per attribute.
• More declared fields ≠ more representational dimensions (geometry): effective rank spans 1.5 (charactereval2) to 81.9 (personahub); MatrAIx's 1,290 categorical attributes collapse to ~6 effective directions.
• Linear persona mixtures degrade with more history; interaction modeling converts facts into capacity (LaMP, 400 users): mean-pooling retrieval MRR peaks at k≈4 (0.263) then declines (0.196 at k=8), while learned attention pooling rises to 0.637 at k=16 — a 2.7× gap. Recency weighting barely helps (0.255 at k=32). Negation cosines 0.70–0.85: text embedders barely separate "likes X" from "dislikes X", so pools need interaction-capable composition.
• Provenance contamination is geometrically exact: personahub and opencharacter-seed share identical geometry (effective rank 81.9, identical EVR, centroid similarity 1.000).
• Dimension scaling does not rescue naive pooling (capacity v2): mean-pooling MRR stays in a flat 0.18–0.26 band from 384 → 1024 dims — the bottleneck is the composition function, not container size. Learned attention dominates at every dimension, and 384-dim MiniLM + attention (0.637 at k=16) beats 1024-dim bge-large + attention (0.498 at k=32). Unsupervised multi-vector centroids (ColBERT-style MaxSim, M=2–8) beat mean pooling modestly (≤0.29) but sit far below learned attention — data-driven slots without task supervision do not find the discriminative structure. Bigger encoders do not fix negation either (spicy-likes/dislikes cosine 0.85 → 0.87).

Method

extract_personas.py (layer 1) loads each source, normalizes records, computes field statistics, embeds up to 400 texts per source with sentence-transformers/all-MiniLM-L6-v2, and computes within/cross-source cosine similarity and family silhouette (sklearn, cosine, seed 42). All 19 sources loaded with zero failures.

llm_extract.py (layer 2) samples 200 personas per source (3,059 total), renders each artifact (short fields first, raw text appended when budget allows), and maps it onto a 20-field unified attribute schema with Qwen/Qwen3-4B-Instruct-2507 (vLLM 0.31.0, temperature 0) with pydantic validation. 354/3,059 (11.6%) outputs failed parsing and are excluded (logged in extractions.jsonl with parse_ok: false); 2,705 records feed the presence-rate tables. Extraction is mapping, not generation: no persona content is invented, and strongest_evidence anchors every non-null attribute to a verbatim quote.

Limitations

• text is truncated to 2,000 chars, capping structured-profile and history lengths.
• The embedding model is English-centric; Japanese text (japan-synthetic-personas) yields near-zero cross-source similarities (~−0.05 to 0.10) — treat those cells as artifacts.
• Small sources are sampled to availability (LaMP: 50 users, Generative Agents: 15 agents, PersonaGym: 100 personas, chub-popular-legacy: 159 cards).
• Sampling caps: 5,000/source for large corpora (Nemotron, PersonaHub, OpenCharacter, chub).

Citation

If you use this corpus, cite the survey "What is a persona representation?" and the source papers above (arXiv ids in source_summary.json).