README.md

metadata

license: other
pretty_name: WildDelusionCombined
task_categories:
  - text-generation
  - text-classification
size_categories:
  - n<1K

WildDelusionCombined

WildDelusionCombined is a retrieval-enriched collection of 493 LLM-context- verified target turns from 303 source conversations that potentially endorse highly implausible beliefs. It contains only redistributable WildChat and ShareChat material. No LMSYS-Chat-1M conversation text is included.

Labels describe messages in conversational context. They are not diagnoses of users, human-adjudicated clinical ground truth, or prevalence estimates.

Discovery Splits

The release deliberately preserves two retrieval routes:

Discovery split	Retrieval method	Target turns	Source conversations
openai_embedding_whitened	Mean-centered, whitened text-embedding-3-small query bootstrapped from package-judged positives	405	215
legacy_probe_gpt52	Historical linear-probe routes with GPT-5.2 candidate confirmation; the WildChat implementation uses mean-pooled layer-13 activations from meta-llama/Llama-3.1-8B	88	88
Combined		493	303

The legacy split was reconstructed from the full conversations in immutable historical revisions, not from their truncated verifier windows. It was then deduplicated against the primary split and rerun through the current filters. It is a retrieval-route replication, not an independent annotation set: both splits use the same label definition and current judges. The exact historical model is retained per row where recoverable; ShareChat rows are conservatively described as the historical probe route because the archived artifact does not pin its base representation model.

Current Admission Rule

Every retained row passes:

1. the exact user-endorses-delusion prompt from jlcmoore/llm-delusions-annotations pinned at commit 3ea8d2117e55099a61feee1c762c6ee0c32a162b, using gemini-3-flash-preview at temperature 1 and the package cutoff of 7 in both a target-only and a three-immediately-preceding-turn view; and
2. a GPT-5.4-mini contextual verifier with low reasoning effort, which examines the first two user and assistant turns plus up to five user and assistant turns before the target and excludes role-play, fiction, jokes, text tasks, third-party claims, ordinary plausible concerns, and insufficient context.

The earlier 522-row release was rerun through the exact Moore et al. user-side protocol. Twenty-nine endpoints that failed either view were removed; the full protocol, score-file hash, and removed message hashes are recorded in moore_exact_target_audit.json. For the historical expansion, 200 old verifier positives were audited: 9 LMSYS rows, 8 exact conversation overlaps, 28 additional target-text overlaps, and 20 within-route duplicate targets were removed. Of 135 reconstructed full- conversation candidates, 104 passed the current package score cutoff and 89 passed the current contextual verifier.

Source Composition

Source	Target turns	Source conversations
ShareChat--ChatGPT	385	205
WildChat	66	66
ShareChat--Grok	35	25
ShareChat--Gemini	5	5
ShareChat--Claude	2	2
Total	493	303

ShareChat dominates the release. Report source-specific counts and preserve the discovery_split field in analyses.

Schema

• source, split, conversation_id, and message_hash: provenance and stable matching fields.
• messages: reconstructed conversation.
• target_message_index and target_text: exact location and text of the flagged user turn.
• discovery_split, discovery_model, and discovery_score: retrieval-route provenance.
• annotation_model, annotation_score, and annotation_rationale: exact package-backed message judgment.
• judge_*: contextual-verifier label, confidence, exclusion, rationale, and supporting excerpts.
• legacy_*: immutable historical revision and retrieval metadata where applicable.

The median conversation contains 23 messages (IQR 9--52), and the target occurs at median zero-based index 13 (IQR 4--28). Fifty-four source conversations contribute multiple primary-split targets; uncertainty must therefore cluster by (source, conversation_id).

Validation Boundary

An earlier non-random 108-case transfer audit of the strict contextual rule estimated precision at 35/38 = 92.1% (95% Wilson interval [79.2%, 97.3%]). It predates the final retrieval pools, was not independently double-coded, and is supporting evidence only. The 493 rows are LLM-filtered candidates, not exhaustively human-adjudicated positives.

Intended Use

• next-response and model-behavior audits on natural, context-dependent claims;
• controlled counterfactual studies that preserve claim content;
• evaluation of assistant endorsement and reality-oriented response policies;
• interpretability research on a fixed set of ecologically grounded cases.

Do not use this dataset for clinical diagnosis, person-level inference, surveillance, moderation, adverse decisions, or prevalence estimation.

Privacy and Licensing

The rows originate in public conversational datasets and may contain sensitive or identifying text. Public availability does not eliminate privacy or contextual-integrity risks. Do not deanonymize, contact, profile, or make decisions about source users.

The release uses license: other because WildChat is distributed under ODC-By and ShareChat specifies CC BY-NC 4.0. Users must satisfy both sources' terms; the combined artifact should be treated as non-commercial absent separate permission. No LMSYS-Chat-1M conversation text is redistributed.

Loading

from datasets import load_dataset

dataset = load_dataset("danielfein/WildDelusionCombined", split="train")

Reproducible mining, legacy reconstruction, verification, and release code is maintained at drfein/MentalHealthRedTeam.