Datasets:
File size: 6,206 Bytes
29d94b3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 | # Competitive position — where we stand vs Google / DeepL / NLLB
**Honest snapshot** (2026-07-21).
**Law:** \(S=K(T_1+T_2+T_3)\) pin **D1D38A** (FSOT; competitors have no equivalent).
---
## Two different games
| Game | What “accuracy” means | Who owns the public narrative |
|------|----------------------|-------------------------------|
| **A. Form→gloss inventory** | Word/lemma → meaning under densify | **PFLT is very strong** |
| **B. Sentence fluency (MT)** | Full sentence → fluent EN (BLEU/chrF/COMET) | **Google / DeepL / NLLB** |
We are **competitive / leading on A**.
We are **climbing toward B**, not at parity yet.
---
## Our numbers (measured in-repo)
### A — Form→gloss solidify (catalog)
| Metric | PFLT |
|--------|------|
| Languages | **113** |
| OPEN / PRODUCT form→gloss | **~99.99%** |
| Thin langs (n<50) | **0** |
| Breadth vs Google ~249 | **~45%** of lang count |
| Breadth vs NLLB ~200 | **~56%** |
| vs DeepL ~30–100 langs | **At/above DeepL scale on count** |
This bar is **not** what Google advertises — but it is real offline inventory depth (incl. classical/historical).
### B — Sentence fluency (dual student: densify + neural)
**B1 — Chat/Tatoeba domain** (16 langs; latest SOTA push 2026-07-21):
| Path | Metric | PFLT | Note |
|------|--------|-----:|------|
| Product densify + residual TM | BLEU-4 / B1 / U-F1 | **~83 / ~98 / ~92** | Product ceiling (includes chat residual templates) |
| **Neural open-set** (best of opus/NLLB/mul per lang, ≤200/lang) | mean sacreBLEU | **~50.2** | Fair chat bar — **past staged 45** |
| Hybrid oracle densify\|neural | mean sacreBLEU | **~53.6** | Upper bound for router |
Per-lang neural peaks: it ~68, es/pt ~61, de/ru ~57–58; thin: hi ~55; CJK gap: ja ~37, zh ~32; classical la densify still wins (~48 product).
**B2 — News domain (public WMT14 de→en test, n=3003)** — harder bar:
| Metric | PFLT densify-only | PFLT neural student (local) | DeepL-class band |
|--------|------------------:|----------------------------:|-----------------:|
| **sacreBLEU** | **~0.4** | **~33.9** (opus-mt-de-en beams=5; NLLB-600M ~33.4) | Often **~40–48+** |
| **chrF** | low | **~60** | Higher |
| Gap to staged mid bar (40) | — | **~6.1** | — |
| Gap to stretch (48) | — | **~14.1** | — |
Domain split: **chat neural is past mid bar** (~50 mean sacre); **news still ~6 BLEU short** of DeepL-class mid. Densify alone does not transfer to news — hybrid router is the product path.
\*Literature ranges, not same-day FLORES A/B. FLORES **data files still 403** (README ok; parquet gated — re-check Hub “Access granted”).
### Law / uniqueness (no competitor equivalent)
| Metric | PFLT |
|--------|------|
| Live pin D1D38A | **Yes** |
| Offline classical + visual path | **Yes** |
| Densify without rewriting law | **Yes** |
---
## Are we competitive at all?
| Track | Competitive? | One-line |
|-------|--------------|----------|
| Offline form→gloss multi-lang | **Yes — strong** | Near-ceiling on our catalog |
| Catalog breadth | **Mid** | 113 of ~200–249; beating DeepL *count* band |
| Classical / dead / hieroglyph | **Yes — unique** | Weak/absent in consumer MT |
| Chat sentence (neural open-set) | **Yes — mid/high open MT** | Mean sacre **~50** (Tatoeba-style) |
| News sentence (WMT de-en) | **Near mid open MT** | **~34** sacre; **~6** to DeepL mid bar |
| Top DeepL commercial SOTA | **Not claimed** | Need +6 to +14 sacre on news |
| Cloud UX / latency | N/A | Different product (local Ada) |
| FSOT law grounding | **Category of one** | — |
**Short answer:**
Yes on **inventory, classical, offline FSOT**, and **chat neural open-set**.
**Almost mid-parity** on news WMT (~34 vs ~40 staged); **not** full DeepL stretch (~48).
Hybrid densify\|neural is the product route (oracle mean chat sacre **~54**).
---
## How far to go (gap to “parity”)
Define stages (honest):
| Stage | Domain bar | Where we are |
|-------|------------|--------------|
| **S0** Dict word map | <5 BLEU | Passed |
| **S1** Phrase densify | ~10–20 | Passed |
| **S2** Strong chat content | ~25–35 BLEU | **Passed** (neural chat mean sacre ~50) |
| **S3** Mid neural news / good EU | ~35–45 sacre WMT | **~34 HERE** — gap **~6** to 40 |
| **S4** Strong neural / DeepL-class | ~45–55+ sacre | Gap **~14** to 48 stretch |
### What must rise for S3–S4
| Lever | Why | Status |
|-------|-----|--------|
| Hybrid router densify\|neural | Product best-of path | **Oracle measured** (~54 chat mean); ship in Ada |
| Word order / T3 (esp. ja/zh) | BP kills densify BLEU despite B1 | **Route CJK to neural** (ja/zh already) |
| FLORES / WMT **same-file** eval | Comparable claim | FLORES data still gated; WMT14 **scored** |
| Neural student (NLLB/opus-mt) | Fluency generators | **Done multi-system**; optional 1.3B / finetune |
| Domain match (news vs chat) | DeepL trains news/web | Need more WMT densify or student finetune |
---
## FLORES status
- Account can read **README** of `facebook/flores`.
- **Data parquet still returns 403** (gated) — accept may be pending, wrong dataset page, or needs re-login/approval.
- Action: open https://huggingface.co/datasets/facebook/flores while logged in → confirm **Access granted** (not just license text).
- Until then we report **Tatoeba product densify** + **WMT** as public bars, not FLORES.
---
## Bottom line
| Question | Answer |
|----------|--------|
| Competitive on form→gloss / offline catalog? | **Yes** |
| Competitive on chat open-set neural? | **Yes — mid/high open MT** (~50 mean sacre) |
| Competitive on Google/DeepL **news** fluency? | **Not parity; S3** (~34 WMT sacre; ~6 to mid bar) |
| How far on news sacreBLEU? | **~6 pts** to staged 40; **~14 pts** to stretch 48 |
| Unique advantage? | **FSOT law + classical + offline densify + hybrid student path** |
**Public storefront:** HF `dappalumbo91/pflt-fsot` · Kaggle `damianpalumbo/pflt-fsot-benchmarks` · GitHub tag/docs `RELEASE_v0.2.0`.
We do **not** claim DeepL parity. We **do** claim a real, measured climb under fixed FSOT law, with high content-level scores and mid full-sentence BLEU.
|