BENCHMARK (v3.0 - STS officiel avec IC bootstrap):
- Use official STS Benchmark (stsb_multi_mt fr, 8628 pairs)
- Add bootstrap confidence intervals (BCa, 95%, 1000 iterations)
- Tian-Dao 20D: Spearman +0.016 [IC: -0.047, +0.042] (non-significant)
- DistilBERT: Spearman +0.775 [IC: +0.777, +0.830] (reference)
- Timestamped reports only (no more duplicates without timestamp)
- Add datasets dependency to requirements.txt
DOCUMENTATION (README.md):
- Add 'Theoretical Foundations' section (bilingual EN/ZH)
- Frame Tian-Dao 20D as signature translation system
- Reference genetic code 64→20 Clifford invariant
- Cite Peter Rowlands & Vanessa Hill (Zero to Infinity, Ch.19)
- Reference DOI: 10.5281/zenodo.19540508
- Update benchmark table with real STS scores + IC
- Clarify Tian-Dao is NOT a semantic model (orthogonal to STS)
- Add topological benchmark section (global score 0.852)
CLEANUP:
- Remove BENCHMARK_REPORT.md (replaced by timestamped version)
- Remove BENCHMARK_RESULTS.json (replaced…
fix: correct indentation and complete encode() method in benchmark
Votre Nom
1516597
- Fix TianDaoEncoder20D class indentation (methods were outside class)
- Complete encode() method with full polarity signature logic
- Align with app.py text_to_embedding() approach
- Use 20 pentad triplets from complexity.pdf Table 1
- Each dimension = one attractor's polarity signature
- Add deterministic noise via np.random.default_rng(hash_val)
- Clip values to [-1, 1] range
fix: align benchmark encoder with app.py 20D approach
Votre Nom
0501ce6
BENCHMARK ALIGNMENT:
- Replace 'Bag of Attractors' approach with 'Polarity signature of triplets'
- Use same 20 pentad triplets as app.py (Table 1 from complexity.pdf)
- Same polarity formula: (n_positive - n_negative) / 3
- Same deterministic noise via np.random.default_rng(hash_val)
- Same clipping to [-1, 1]
COHERENCE:
- Benchmark now uses IDENTICAL logic to text_to_embedding() in app.py
- Both produce 20D embeddings with same structure
- Each dimension = one attractor's polarity signature
- Eliminates discrepancy between app and benchmark
DOCUMENTATION:
- Update TianDaoEncoder20D docstring to reflect new approach
- Clarify that benchmark and app use same encoding logic
