| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| task_categories: |
| - text-classification |
| tags: |
| - safety |
| - content-moderation |
| - slang |
| - youth-language |
| - benchmark |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: Gen_Alpha_Benchmark_v7_5.csv |
| - config_name: baseline_v7 |
| data_files: |
| - split: train |
| path: "Gen Alpha Baseline - V7.csv" |
| --- |
| |
| # GenAlphaBench |
|
|
| A safety-annotated benchmark of 239 Generation Alpha (born 2010-2024) expressions for |
| evaluating LLM comprehension and safety assessment of youth digital language. Each |
| expression is annotated across a six-category risk taxonomy (R0 neutral through R5 |
| cultural variant), platform contexts, community origin, temporal metadata, and usage |
| examples, with per-expression source documentation for origin, meaning, and evolution |
| claims. |
|
|
| **Canonical file: `Gen_Alpha_Benchmark_v7_5.csv`** (see `CHANGELOG.md` for the version |
| history). Earlier versions are retained: `Gen_Alpha_Benchmark_v7_4_2.csv` and |
| `Gen_Alpha_Benchmark_v7_4_1.csv` (intermediate provenance releases) and |
| `Gen Alpha Baseline - V7.csv`, the version the paper's model evaluations were run |
| against, kept for reproducibility. The baseline differs from the canonical file at |
| ids 220 and 232, where two expressions were substituted, and lacks the provenance |
| columns. |
|
|
| ## Provenance columns |
|
|
| - `origin_sources` — documentary sources supporting each community-origin label |
| (format `type:URL | type:URL`, two independent sources required); full audit trail |
| with per-source evidence quotes in `origin_sources_audit.csv` |
| - `attribution_basis` — `documentary` (two sources), `documentary_single` (one), or |
| `informant` (native-speaker attestation by the Gen Alpha member of the research team) |
| - `meaning_sources` / `meaning_basis` — a public source defining the expression |
| consistently with `basic_meaning`; basis is `informant_cohort_documentary` or |
| `informant_cohort` (native-speaker authorship validated by the 24-member peer cohort) |
| - `evolution_sources` / `evolution_basis` — dated documentation of the meaning shift |
| described in `evolution_notes`; `informant_documentary` or `informant` |
|
|
| Coverage: 234/239 meanings and 184/239 evolution notes carry documentary sources; |
| the remainder are explicitly informant-based, concentrated among post-2024 expressions |
| that no dictionary or reference site has yet recorded — absence of documentation for |
| recent coinages is part of what the benchmark measures. In v7.4.1, nine origin labels |
| were re-coded to `none` after the documentary audit; in v7.4.2, sixteen expressions |
| whose audited sources attribute origin to Black queer, drag, or ballroom culture were |
| re-coded to the joint `AAVE/LGBTQ+` value. |
|
|
| ## Content notice |
|
|
| The dataset documents coded slurs, suicide-baiting language, sexual euphemisms, and |
| references to violence, included for the purpose of evaluating detection capabilities. |
| Community-origin labels exist for bias auditing only. Not intended for: training |
| harmful-content generators, differential moderation by community origin, or |
| surveillance of youth communication. |
|
|
| ## Collection |
|
|
| Expressions were identified through observation of publicly circulating vocabulary and |
| verified against public documentation. The dataset contains no user posts, quotations, |
| usernames, or personal data; usage examples are composed by the research team. Risk |
| labels were validated by a cohort of 24 Gen Alpha participants (ages 11-14, parental |
| consent, kappa = 0.81); label disagreements were resolved by expert review. |
|
|
| ## Evaluation code and leaderboard |
|
|
| https://github.com/SystemTwoAI/GenAlphaSlang — verbatim prompts, scoring pipeline, |
| validation scripts, and a public model leaderboard under a pinned LLM-judge protocol. |
|
|
