---
license: other
language:
- en
task_categories:
- text-generation
pretty_name: WildDelusion Benchmark
size_categories:
- n<1K
---

# WildDelusion Benchmark

WildDelusion Benchmark is the minimal, peer-facing release of the final
Gemini-filtered WildDelusion cohort. It contains 493 verified user target turns
from 303 naturally occurring, publicly shared human-chatbot conversations.

Each row includes the complete reconstructed conversation, including every
available assistant turn. Roles are normalized to `user` and `assistant`, and
`target_message_index` is a zero-based index into `messages` that points exactly
to `target_text`.

## Data

The dataset intentionally has only six columns:

| Column | Description |
|---|---|
| `example_id` | Stable SHA-256 identifier for the verified target turn. |
| `source` | Source corpus/model-family route. |
| `conversation_id` | Source conversation identifier; use with `source` for clustering. |
| `messages` | Complete ordered conversation as `{role, content}` records. |
| `target_message_index` | Zero-based location of the verified user target in `messages`. |
| `target_text` | Exact text at `messages[target_message_index]`. |

For response evaluation, provide the conversation only through the target user
turn. Later messages are retained for longitudinal research but must not be
included in the model input:

```python
from datasets import load_dataset

dataset = load_dataset("danielfein/WildDelusionBenchmark", split="train")
row = dataset[0]
model_input = row["messages"][: row["target_message_index"] + 1]
```

## Selection

This release is an exact projection of the 493 targets retained in
`danielfein/WildDelusionCombined` revision
`4749e769261c6b068da4a42aa0f91ff4903c2985`. Admission required a score of at
least 7 under the exact Moore et al. `user-endorses-delusion` rubric using
`gemini-3-flash-preview` in both target-only and three-preceding-turn views.
Twenty-nine of 522 pre-filter targets were excluded.

The conversations were independently rehydrated from pinned ShareChat and
WildChat revisions. The release contains 24,170 user turns and 24,141 assistant
turns. All 493 transcripts contain assistant messages; 425 targets have at least
one assistant turn before the target. One conversation includes four assistant
image-generation outputs whose text was unavailable in the source export; their
positions are preserved with explicit non-text placeholders.

See `manifest.json` for pinned revisions and `integrity_report.json` for the
machine-readable audit.

## Scope and limitations

- Inclusion is a rubric-based content annotation, not a clinical diagnosis of a
  user or a claim about the user's mental state.
- The dataset is retrieval-enriched and must not be used to estimate population
  prevalence.
- Multiple targets may come from the same source conversation. Split or bootstrap
  by `(source, conversation_id)`, not by row.
- The material can contain sensitive mental-health, self-harm, conspiratorial,
  and personally revealing content. Use appropriate access and review practices.
- Exact source model versions are not consistently available, so `source`
  identifies the corpus/model family rather than a guaranteed model snapshot.

## Licensing

This combined release is non-commercial. ShareChat-derived content is subject to
CC BY-NC 4.0 terms; WildChat-derived content is subject to ODC-By. Users are
responsible for complying with the upstream datasets' terms and privacy guidance.