dappalumbo91/pflt-fsot

Protofluid Language Translator (PFLT) — FSOT-native

Version: 0.2.8 · Snapshot: 2026-07-23
Author: dappalumbo91
Code: github.com/dappalumbo91/protofluid-language-translator
Dataset mirror: dappalumbo91/pflt-fsot-sample

Naming: FSOT_* = our product/ranking · NLLB/OPUS = competitor student generators · DeepL = external bar only.

---

What this is

PFLT is an offline-first translator intelligence surface grounded in the FSOT free-parameter law:

[ S = K(T_1 + T_2 + T_3) ]

Authority pin: D1D38A (SHA-256 prefix of archive-canonical fsot_compute.py).

It is not “another LLM wrapper.” It is:

1. Form→gloss catalog densify under fixed law (113 languages, near-ceiling product/open inventory)
2. FSOT Language Second Brain — form↔sense↔language graph + FSOT pathway edges
3. Neural students (OPUS-MT, NLLB-600M / 1.3B / 3.3B) for full-sentence fluency
4. FSOT ranking over student multi-hyp pools (not from-scratch NMT weights)
5. Ada/SPARK product binary in the GitHub repo (Python = data factory)

Honest claims (read first)

Claim	Status
Google / DeepL commercial news SOTA	Not claimed
Chat-domain open-set MT (Tatoeba-style)	Competitive mid/high open MT (~50 mean sacreBLEU)
WMT14 de→en FSOT product	36.80 (FSOT_pick_hardset) · ~92% of DeepL mid-40
vs NLLB-3.3B single beam	FSOT product edges multi-hyp ranking (36.80 vs 36.74) — students still generate
FSOT_oracle_pool (selection ceiling)	46.12 — headroom is ranking, not more beams alone
Form→gloss multi-lang catalog	Strong / near-ceiling (~99.99% on 113 langs)
FSOT law uniqueness	Category of one (pinned (S=K(T_1+T_2+T_3)))

Product densify BLEU can look extremely high when residual translation-memory templates are installed for known chat — that is a product ceiling, not open-set generalization. Fair open-set bars are the neural student numbers below.

---

Benchmarks (v0.2.8)

Protocol details and JSON: metrics_snapshot.json in this repo.

A — Catalog (form→gloss)

Metric	Value
Language codes	113
OPEN form→gloss	~99.99%
PRODUCT form→gloss	~99.99%
Thin langs (n<50)	0

B1 — Chat / Tatoeba-style (open-set neural)

Best of local students per language, ≤200 sentences/lang, beams=5, sacreBLEU:

Aggregate	sacreBLEU
Mean best (16 langs)	50.19
Hybrid oracle densify|neural	53.58
Staged bar (internal)	45.0

Lang	Best system	sacreBLEU
it	nllb-600M	67.83
es	opus-mt	61.34
pt	nllb-600M	61.29
ru	nllb-600M	57.75
de	opus-mt	57.10
nl	nllb-600M	56.79
fr	nllb-600M	56.54
hi	nllb-600M	54.73
tr	nllb-600M	54.28
ar	opus-mt	53.27
pl	nllb-600M	52.66
he	nllb-600M	47.88
ko	nllb-600M	39.81
ja	opus-mt	36.55
zh	nllb-600M	32.20
la	opus-mt-mul-en	12.95

Latin / some classical pairs still prefer densify product path in hybrid routing.

B2 — News (WMT14 German→English test, n=3003)

System	Owner	sacreBLEU
FSOT_pick_hardset (product)	FSOT	36.80
FSOT_product_gen	FSOT	36.79
NLLB-200-3.3B (beams=8)	Meta student	36.74
NLLB-200-3.3B (beams=5)	Meta student	36.69
NLLB-200-1.3B (beams=5)	Meta student	35.63
OPUS-mt-de-en (beams=5)	Helsinki student	33.88
NLLB-600M distilled	Meta student	33.37
FSOT_oracle_pool	FSOT ceiling	46.12
DeepL-class mid bar	external	~40
Stretch SOTA bar	external	~48
Gap FSOT product → DeepL mid-40		3.2
Gap FSOT product → FSOT_oracle_pool	selection	9.32

See NEWS_DEEN_CACHED.md, LANGUAGE_SECOND_BRAIN.md, metrics_snapshot.json.

Stage ladder

Stage	Meaning	Status
S0–S1	Dict / phrase densify	Passed
S2	Strong chat content	Passed (neural chat ~50)
S3	Mid neural news (~35–45 sacre)	~34 — HERE
S4	DeepL-class stretch (~45–55+)	Gap ~14 sacre on WMT

---

Architecture (FSOT)

Law (fixed)  S = K(T1+T2+T3)  pin D1D38A
     │
     ├─ Densify surface   form→gloss, phrases, classical
     ├─ Neural students   OPUS-MT / mul-en / NLLB (decode only)
     └─ Hybrid router     domain/length gate (oracle measured)

Students never rewrite the law scalar. Densify grows the surface only.

---

What ships on the Hub

File	Description
README.md	This model card
metrics_snapshot.json	Machine-readable benchmarks v0.2.0
sample_densify.tsv	Small densify sample (demo)
app.py	Gradio form→gloss + FSOT panel demo
requirements.txt	Demo deps

Not on the Hub (by design): multi-GB densify/gold packs, Ada binary, full Kaikki/Tatoeba dumps (stay on local D:).

---

Quickstart

Local Gradio demo

git clone https://github.com/dappalumbo91/protofluid-language-translator.git
cd protofluid-language-translator/huggingface_pflt
pip install -r requirements.txt
python app.py

Full product (Ada)

cd pflt-Ada
alr build
.\bin\pflt_main.exe eval-product
.\bin\pflt_main.exe archive   # live D1D38A pin

Reproduce sentence benchmarks (local models required)

cd pflt-Ada
python -u m6_sota_push.py
# writes reports/M6_SOTA_PUSH.md + m6_sota_push_report.json

Models expected under a local cache (example):

• Helsinki-NLP/opus-mt-{es,de,fr,ru,zh,ja,ar}-en
• Helsinki-NLP/opus-mt-mul-en
• facebook/nllb-200-distilled-600M

---

Remaining path to SOTA (parked for later)

1. Ship hybrid router in Ada product decode
2. Optional WMT student finetune (still densify under law)
3. FLORES when Hub data access unlocks
4. Optional NLLB 1.3B / larger beams for news
5. CJK: keep neural path (SPM), densify for inventory only

---

Citation / links

@software{pflt_fsot_2026,
  title  = {Protofluid Language Translator (PFLT) under FSOT},
  author = {Palumbo, Damian},
  year   = {2026},
  url    = {https://github.com/dappalumbo91/protofluid-language-translator},
  note   = {Law pin D1D38A; version 0.2.0 benchmarks}
}

• GitHub: https://github.com/dappalumbo91/protofluid-language-translator
• Competitive position: docs/COMPETITIVE_POSITION.md
• SOTA report: docs/M6_SOTA_PUSH.md
• Kaggle benchmarks pack: search pflt-fsot-benchmarks under user damianpalumbo

License

Apache-2.0 for code. Sample densify rows derived from Wiktionary-class sources — respect upstream CC-BY-SA/GFDL as applicable. Large third-party corpora (Tatoeba, WMT, OPUS, NLLB weights) remain under their original licenses and are not redistributed here.