Upload folder using huggingface_hub

README.md

@@ -16,48 +16,49 @@ configs:
 
 # Hebrew Semantic Neighbor Pairs
 
-7,417 Strong's-number pairs of Biblical Hebrew words judged to be synonyms or same tight semantic
-concept, at a publication-confidence level reached through two independent verification methods rather
-than any single signal. CC0 — no restricted sources.
+7,977 Strong's-number pairs of Biblical Hebrew words judged to be synonyms or same tight semantic
+concept, at a publication-confidence level reached through three independent verification methods
+rather than any single signal. CC0 — no restricted sources.
 
 ## Method
 
-This dataset is the *confident subset* of a larger semantic-neighbor pipeline that combines nine
+This dataset is the *confident subset* of a larger semantic-neighbor pipeline that combines ten
 independent signals: distributional embeddings (BEREL), LXX cross-testament co-rendering, gloss
 overlap, an LLM scholarly-prior pass, Brown-Driver-Briggs (1906, public domain) etymological roots,
 T'OMIM poetic-parallelism pairs, Hebrew WordNet, BHSA syntactic structure (coordination/apposition via
-Context-Fabric), and a cross-lingual/Wiktionary-root corroboration tier. Individually, these signals
-range from weak (~35%) to strong (~90%).
+Context-Fabric), a cross-lingual/Wiktionary-root corroboration tier, and candidates from Radak's Sefer
+HaShorashim (a 13th-century Public Domain medieval Hebrew root dictionary, via Sefaria). Individually,
+these signals range from weak (~35%) to strong (~90%).
 
-Rather than trust any single signal's raw output, a pair is published here only if it passes one of two
-independent confidence gates:
+Rather than trust any single signal's raw output, a pair is published here only if it passes one of
+three independent confidence gates:
 
-- **`cross_signal`** — asserted by >= 2 methodologically independent signal *families*
-  (distributional / etymological / structural / lexical-external / scholarly-LLM / cross-lingual —
-  see `gate` column). 3,225 pairs.
-- **`llm_verified`** — for pairs that only ever had ONE signal behind them, a dedicated LLM pass
-  (Claude Haiku, strict "are these genuinely synonyms in Biblical Hebrew usage" prompt) judged each one
-  individually. Only "yes" verdicts are included here — pairs judged "no" or "unsure" are not published.
-  4,192 pairs.
+- **`cross_signal`** — asserted by >= 2 methodologically independent signal families (see the
+  `gate` column). 70.2% SDBH `core`-agreement on its own.
+- **`llm_verified`** — pairs that only ever had ONE signal behind them, individually judged by an LLM
+  (Claude Haiku, strict "are these genuinely synonyms in Biblical Hebrew usage" prompt). Only "yes"
+  verdicts included. 69.5% on its own.
+- **`sefer_hashorashim_verified`** — candidates from Sefer HaShorashim (words Radak discusses together
+  within one root entry), also LLM-verified the same way. 62.9% on its own (its unique contribution
+  after removing overlap with the other two gates).
 
-No pair passes both gates by construction (the LLM pass only ran on pairs that had exactly one signal,
-by design — so the two gates cover disjoint input, not double-confirmation).
+1,241 pairs pass more than one gate — that subset scores 78.4%, the highest-confidence layer in the
+dataset (see the `gate` column, "+"-joined for multi-gate pairs).
 
 ## Columns
 
 | column | meaning |
 |---|---|
 | `strong_a` / `strong_b` | Hebrew Strong's numbers (`H####`) |
-| `gate` | `cross_signal` or `llm_verified` — which confidence check this pair passed |
+| `gate` | which confidence check(s) this pair passed — `cross_signal`, `llm_verified`, `sefer_hashorashim_verified`, or a "+"-joined combination |
 | `n_families` | how many independent signal families asserted this pair (0 if only LLM-verified) |
-| `llm_verdict` | `yes` if LLM-verified; blank if published via cross-signal agreement alone |
+| `llm_verdict` | `yes` if LLM-verified (either LLM gate); blank if published via cross-signal agreement alone |
 
 ## Validation caveat — read before treating the percentages as ground truth
 
-Every accuracy number associated with this pipeline (73.4% for `cross_signal`, 69.6% for
-`llm_verified`) is measured as **agreement with SDBH's own domain codes** (the Semantic Dictionary of
-Biblical Hebrew, © United Bible Societies) as an internal proxy — not an independent audit, and not a
-claim of "correctness." Two things follow:
+Every accuracy number associated with this pipeline is measured as **agreement with SDBH's own domain
+codes** (the Semantic Dictionary of Biblical Hebrew, © United Bible Societies) as an internal proxy —
+not an independent audit, and not a claim of "correctness." Two things follow:
 
 1. SDBH is a scholarly taxonomy, not exhaustive ground truth — a real synonym pair can fail to share an
    SDBH domain code simply because SDBH's own domain boundaries split it, and (separately) SDBH data
@@ -70,8 +71,9 @@ claim of "correctness." Two things follow:
 
 CC0-1.0. Built from: BDB (1906, public domain), T'OMIM (CC BY 4.0), Hebrew WordNet (Ordan & Wintner,
 permissive), Wiktionary (CC BY-SA), BHSA/Context-Fabric (public-domain WLC + CC BY 4.0 ETCBC
-annotation), and an LLM verification pass (Claude Haiku) — none of these require restricted-use
-attribution beyond what's already satisfied by citing this dataset.
+annotation), Sefer HaShorashim (Radak, c.1185-1235 CE, Public Domain, via Sefaria), and an LLM
+verification pass (Claude Haiku) — none of these require restricted-use attribution beyond what's
+already satisfied by citing this dataset.
 
 ## Provenance