Upload 4 files

CHANGELOG.md

@@ -1,3 +1,30 @@
+# Dataset v7.5
+
+- Added four columns extending source documentation from origin labels to the meaning
+  and evolution fields: `meaning_sources` (a public source defining/glossing the
+  expression consistently with basic_meaning), `meaning_basis`
+  (informant_cohort_documentary | informant_cohort), `evolution_sources` (dated
+  documentation of the meaning shift described in evolution_notes), and
+  `evolution_basis` (informant_documentary | informant).
+- Coverage: 234/239 expressions carry a meaning source; 184/239 carry a dated
+  evolution source. The remainder are marked informant-basis: meanings established
+  by native-speaker authorship and validated by the peer cohort, evolution notes by
+  native-speaker observation. Absence of documentation for recent expressions is
+  expected and is part of what the benchmark measures.
+- Sources were located with search tooling and verified by the authors; the audit
+  trail is in meaning_evolution_verification_checklist.csv pending sign-off.
+- No changes to any pre-existing field.
+
+# Dataset v7.4.2
+
+- Multi-label community-origin conversion: 16 expressions whose audited sources attribute
+  origin to Black queer/drag/ballroom culture re-coded from single-label (AAVE or LGBTQ+)
+  to the joint AAVE/LGBTQ+ value, joining the 2 items already so coded. The re-coding is
+  evidence-driven: only rows whose recorded origin_sources name both lineages were changed.
+  Deliberately unchanged: "twink" (documented US gay slang, 1960s) and "fruity" (Polari
+  lineage), whose records do not indicate AAVE origin.
+- No other fields changed from v7.4.1.
+
 # Dataset v7.4.1 — 2026-07-23
 
 - Added `origin_sources`: per-expression documentary sources supporting community-origin labels (format `type:URL | type:URL`). 60 expressions carry two independent sources; 7 carry one.

README.md

@@ -16,7 +16,7 @@ configs:
 - config_name: default
   data_files:
   - split: train
-    path: Gen_Alpha_Benchmark_v7_4_1.csv
+    path: Gen_Alpha_Benchmark_v7_5.csv
 - config_name: baseline_v7
   data_files:
   - split: train
@@ -25,28 +25,41 @@ configs:
 
 # GenAlphaBench
 
-A safety-annotated benchmark of 239 Generation Alpha (born 2010–2024) expressions for
+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 24 fields: a six-category risk taxonomy (R0 neutral
-through R5 cultural variant), platform contexts, community origin, temporal metadata
-(origin, mainstream adoption, decline, staleness), and usage examples.
+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_4_1.csv`** (v7.4.1 — see `CHANGELOG.md`).
+**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.
 
-`Gen Alpha Baseline - V7.csv` is retained for provenance: it is the version the paper's
-model evaluations were run against. The two files differ at ids 220 and 232, where
-v7.4.1 substitutes two expressions, and in v7.4.1's two added columns:
+## Provenance columns
 
 - `origin_sources` — documentary sources supporting each community-origin label
-  (format `type:URL | type:URL`); the full audit trail with per-source evidence notes
-  is in `origin_sources_audit.csv`
-- `attribution_basis` — `documentary` (two independent sources), `documentary_single`
-  (one), or `informant` (native-speaker attestation by the Gen Alpha member of the
-  research team)
+  (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`
 
-In v7.4.1, nine community-origin labels from earlier versions were re-coded to `none`
-after a documentary audit found no supporting attribution or documented alternative
-origins; prior labels are preserved in the audit file.
+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
 
@@ -59,11 +72,10 @@ surveillance of youth communication.
 ## Collection
 
 Expressions were identified through observation of publicly circulating vocabulary and
-verified against public documentation (Know Your Meme, slang lexicography, dated news
-coverage, Google Trends). 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,
-κ = 0.81); label disagreements were resolved by expert review.
+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