pflt-fsot-sample / docs /FSOT_FLUENCY_GAP.md
dappalumbo91's picture
v0.2.1: FULL verification source (not metrics-only)
29d94b3 verified
|
Raw
History Blame Contribute Delete
2.78 kB

FSOT fluency gap solver — compete via mathematics of densify

Built: 2026-07-21T18:27:20.793197+00:00

Law (fixed — not fitted to BLEU)

  • Formula: S=K(T1+T2+T3) pin D1D38A
  • Linguistic panel: S=0.651324761885 T1=0.549955 T2=1.000000 T3=1.707063e-17
  • Densify students cannot rewrite K or seeds.

Surface mapping (competitiveness as FSOT-shaped S)

Component Surface meaning Metric drivers
T1_surf Lexicon / morph base coverage × U-F1
T2_surf Linear phrase map BLEU-1 × bigram retention
T3_surf Order / fluency valve BP × chrF
S_surf K·(T1+T2+T3) staged competitor bar

Staged bar S_surf ≈ 0.72008 (T1=0.650 T2=0.650 T3=0.414)

Before → after densify

Metric Before After Δ
BLEU-4 2.2 18.82 16.62
BLEU-1 45.87 86.86 40.99
U-F1 34.81 77.78 42.97
chrF 25.07 61.47 36.4
Coverage 91.3 100.0

Densify inject: +0 unigrams, +0 bigrams/templates (product path).

Per-lang priority (ΔS densify budget)

Lang ΔS Priority Budget B1 after BLEU after
zh 0.6562 T1 1.2848 100.0 12.15
ja 0.6457 T1 1.2642 99.61 10.82
ko 0.4289 T2 0.8397 95.88 28.47
hi 0.3787 T2 0.7414 42.79 3.78
tr 0.3688 T2 0.7221 95.2 22.16
ar 0.3533 T2 0.6917 92.49 26.81
de 0.3389 T2 0.6636 81.11 17.05
he 0.3116 T2 0.6101 96.37 23.88
ru 0.3014 T2 0.5900 93.09 21.08
la 0.3010 T2 0.5894 97.5 22.15
fr 0.2944 T2 0.5764 79.65 14.62
pl 0.2858 T2 0.5596 95.44 20.2
es 0.2517 T2 0.4927 80.74 16.25
nl 0.2341 T2 0.4583 81.63 20.99
pt 0.2165 T2 0.4238 81.65 16.79
it 0.2139 T2 0.4188 85.41 21.96

How this is solving toward competitiveness

  1. Measure gap as ΔS_surf, not ad-hoc loss.
  2. Allocate densify by component shortfall (T1/T2/T3) × archive growth.
  3. T3-priority langs get phrase templates + light EN reorder valve.
  4. Law S stays the constitution for converse/cert; fluency is student densify.

Staged bars are intermediate (BLEU-4≈20, B1≈65), not final Google/DeepL SOTA. Solving ΔS_surf with densify raises surface fluency; law S unchanged.

Next FSOT-native levers

  • Raise staged bar toward neural (BLEU-4 30+) once T2/T3 densify saturates
  • NLLB teacher densify for T2 (still student)
  • FLORES eval when Hub access accepted
  • Optional neural student under cert gate — densify only