This dataset contains synthetic sentences that express or describe suicidal ideation and self-harm intent. It is a research and transparency artifact only. By requesting access you confirm that you will use it solely for non-commercial research or teaching, that you will not use it to screen, triage, diagnose, or make any decision about a real person, and that you accept the CC BY-NC-SA 4.0 license and the limitations described in the dataset card.

Suicidality Detection Corpus (multilingual, synthetic)

⚠️ Content warning. This dataset contains sentences that express or describe suicidal ideation, self-harm intent, hopelessness, and coded farewells, in five languages. It exists for research on text-classification data handling. It is not a clinical instrument and must never be used to screen, triage, diagnose, or make any decision about a real person.

What this is

A labelled, multilingual corpus for the binary task "does this sentence carry a suicidality signal?". Each row is a short first-person message labelled suicidality or neutral.

The data is synthetic. These are generated and hand-cleaned clinical-style sentences, not scraped posts from real people: there are no usernames, URLs, or social-media artifacts, and no real individual's crisis text. That is deliberate, and it is why the set can be released at all.

This corpus was the suicidality class of an internal Input-Risk-Scorer training set at Silicon Psyche Labs. On 2026-08-19 the company removed user-suicidality from its product entirely: the production reader was retrained without this class, and no runtime path scores, stores, or alarms on a user's suicidality any more. Reading a user's suicidal state is special-category health data and sits outside what the product measures (the machine's behaviour, read from its output). The corpus is published here, on its own, as a data-handling and transparency artifact, and is not part of any production model.

Schema

Newline-delimited JSON (suicidality_detection_corpus.jsonl); one object per line:

field	type	meaning
`text`	string	the message
`label`	`"suicidality"` \| `"neutral"`	class
`lang`	string \| null	`en`/`it`/`fr`/`es`/`de`, or `null` where the source left it untagged
`split`	`"train"` \| `"test"`	the split the row was drawn from
`id`	string	stable row id

Composition

3,511 rows: 422 suicidality, 3,089 neutral.
Languages: en, it, fr, es, de (plus untagged rows).
The neutral rows include benign, technical, and near-miss text (bereavement, resignation, academic references to death) so a classifier is forced to separate a real signal from surface vocabulary rather than keyword-match.

Intended use

Research and teaching on multilingual crisis-text classification, data documentation, and the false-positive / false-negative trade-off on a high-stakes construct.

Explicitly out of scope

Any real-time screening, triage, or intervention for a real person.
Any claim of clinical validity. Labels are text annotations, not clinical judgements.
Training a product that reads a user's suicidality. The originating company removed that feature by decision; this dataset is a record, not a starting point for reinstating it.

If you or someone you know is in crisis

Please reach a local crisis line. In the US you can call or text 988 (Suicide & Crisis Lifeline). International directories: https://findahelpline.com.

License

CC BY-NC-SA 4.0, with the additional out-of-scope limits above. Access is gated: you accept these terms before downloading.

Citation

@misc{siliconpsyche2026suicidalitycorpus,
  title  = {Suicidality Detection Corpus (multilingual, synthetic)},
  author = {Silicon Psyche Labs},
  year   = {2026},
  note   = {Research artifact; the suicidality class removed from the Silicon Psyche production reader}
}

Downloads last month: 14

Total file size:

569 kB