Skip to content

v0.7.0

Choose a tag to compare

View all tags
@JoshuaRamirez JoshuaRamirez released this 17 Apr 02:51
· 6 commits to main since this release
v0.7.0
e27ae00

[0.7.0] - 2026-04-16

Added

  • BGE-large-en-v1.5 as sole embedding source at native 1024d (ADR-013). Replaces GloVe + Model2Vec paths. Runtime remains numpy-only; BGE is a build-time dependency under a new [build] optional extra (sentence-transformers, torch, wordfreq, nltk).
  • wordfreq for frequency ordering and top-N vocabulary selection — no more GloVe dependency of any kind.
  • FACE_VERNACULAR dictionary — targeted object-level vocabulary for underperforming faces (ethics, axiology, methodology) only. Added at 4x weight in face centroid construction.
  • scripts/expand_pole_synonyms.py — WordNet-based pole-synonym expansion proposals for human review.
  • ADR-013: rationale for BGE as the sole build-time embedding source.

Changed

  • Build script rewritten (~2080 → ~770 lines). Deleted Model2Vec path, face-informed QR+PCA reduction, 400K→15K vocab trimming, OOV expansion, counter-fitting machinery. Single unified vocab.
  • Questions and phrases are now encoded as full sentences through BGE, not as IDF-weighted word-vector averages.
  • face_relevance uses IDF² weighting — rare face-specific vocabulary dominates the weighted average; common tokens no longer drown signal.
  • 20% per-face column centering applied at artifact load — corrects systematic bias where "specific" face centroids (ethics, methodology) accumulated negative bias from generic tokens.
  • Cube-pair contrast dampening disabled: the mechanism punished legitimate co-activations (ethics+axiology on moral rhetoric, epistemology+methodology on scientific texts). Re-enable only if evidence supports it.
  • docs/specs/semantic-bridge-algorithms.md updated for BGE pipeline.

Measurement

  • Benchmark: scripts/benchmark_8texts.py score improved from 13/20 → 17/20 (+31%).
  • Aesthetics moved from rank #10 to #1 on Aristotle's Poetics; methodology now hits top-6 on Newton; phenomenology-dominates-everything ceiling broken.

Removed

  • GloVe download/loading code paths.
  • Model2Vec integration and PCA reduction machinery.
  • select_runtime_vocab 400K→15K trimming (vocab is assembled directly from what's needed).
  • expand_vocab_by_question_proximity OOV expansion pass.
  • Counter-fitting SGD (retired pending evidence that contextual embeddings still benefit).
Assets 2
Loading