| # News mid-parity push — measured
|
|
|
| **Built:** 2026-07-22T17:10:39.514108+00:00
|
| **Law pin:** D1D38A
|
|
|
| ## Headline
|
|
|
| | System | sacreBLEU | Gap to 40 |
|
| |--------|----------:|----------:|
|
| | Base opus-mt-de-en | 33.88 | 6.12 |
|
| | NLLB-600M | 33.37 | 6.63 |
|
| | **Product cross-NLL ensemble** | **34.11** | **5.89** |
|
| | FT v3 (early-stop) **test** | 33.41 | 6.59 |
|
| | FT v3 best **val** | 35.76 | (not test) |
|
| | Oracle uF1 ensemble | 37.13 | 2.87 |
|
|
|
| **Mid-40 cleared?** No.
|
|
|
| ## Finetune v3
|
|
|
| - Froze first 3/6 encoder layers, lr=3e-5, train=120k pairs
|
| - Early-stop patience 4 @ every 500 steps
|
| - Best val **35.76** @ step 500; test **33.41** (do not ship over base)
|
| - Saved: Helsinki-NLP__opus-mt-de-en-wmt-ft-v3
|
|
|
| ## Product ensemble
|
|
|
| Cross-model NLL pick: **34.11** (opus 1776 / nllb 1227).
|
| Same as MBR-cross-NLL. Oracle still **37.13** (~3 pts selection headroom).
|
|
|
| ## Chat (prior lever)
|
|
|
| Neural-first hybrid: **48.74** sacre — past mid bar 45.
|
|
|
| ## Honest remaining path to beat news mid-bar
|
|
|
| 1. Learned ensemble / quality estimator (close 34.1 → ~37)
|
| 2. Larger student (NLLB-1.3B) or multi-epoch FT with true WMT val that tracks test
|
| 3. FLORES multi-pair when unlocked
|
|
|
| ## Repos
|
|
|
| Code + this report shipped to GitHub / HF / Kaggle with verify pack.
|
| |
