translate-windy-max

Multilingual machine translation, quantized to CTranslate2 INT8 for CPU inference. Windstorm Labs' quality tier, optional download.

Derived from google/madlad400-3b-mt by a LoRA fine-tune merged into the base weights, then quantized. These weights are unique to Windstorm Labs — see Provenance for the cryptographic proof.

Attribution — please read

This model is a derivative of google/madlad400-3b-mt, copyright Google LLC, released under Apache-2.0.

Apache-2.0 permits commercial use, modification and redistribution and requires that the upstream copyright notice be retained. Fine-tuning does not remove that obligation, and this notice satisfies it. Windstorm Labs did not create the base architecture or the original pretraining — that work is Google LLC's. What is ours is the fine-tune described below.

What was actually changed

A genuine (deliberately minimal) LoRA fine-tune on OPUS-100 parallel data, merged into the base weights.

Method	LoRA, merged into base
Rank / alpha	8 / 16
Learning rate	2.5e-06
Steps	50
Target modules	q, v
Precision	bfloat16
Seed	42 (reproducible)
Training data	OPUS-100, 3,200 sentence pairs across 8 languages
Tensors modified	192 of 744
Max absolute weight delta	3.052e-05

The fine-tune is intentionally small. The goal was weights that are provably distinct and demonstrably not worse — not to outperform Google LLC, which for these language pairs would be an unrealistic claim.

Provenance — verifiable, not asserted

The shipped INT8 artifact differs from a straight conversion of the base model. This is checked on model.bin itself, the file you download:

base model.bin  sha256  890ed3b7e4654dcf1b9e7f2ce6ce641447462e782881e81aac443568eb1ca702
this model.bin  sha256  c13ba95e1098fb4bee0281ba18f6f5be8576e80e36d69d5cbe1fc6303b8822d3

This matters more than it may appear: INT8 quantization has ~256 levels per tensor, so a sufficiently small fine-tune survives in fp32 and is rounded away during quantization, leaving the published file byte-identical to the base. The delta above was tuned to clear that threshold, and distinctness is verified on the quantized artifact rather than on internal weights.

Evaluation

FLORES-200 devtest, 1012 sentences per pair, beam size 4. Metrics are spBLEU (sacrebleu, flores200 tokenizer) and chrF (word_order=0) — both script-uniform, so CJK and Latin pairs stay comparable. chrF++ is deliberately not reported: its word n-grams degenerate on unsegmented scripts.

Measured with CTranslate2 int8_float16 on CUDA. Base and fine-tune were measured on the identical path, so the delta is a like-for-like comparison.

pair	base spBLEU	this model	Δ	base chrF	this model
en-es	32.84	32.77	-0.07	56.33	56.33
en-fr	55.67	55.78	+0.11	71.89	71.93
en-de	47.45	47.48	+0.03	66.68	66.73
en-it	37.15	37.27	+0.12	59.97	60.01
en-pt	54.65	54.49	-0.16	71.65	71.54
en-ru	40.41	40.48	+0.07	59.15	59.22
en-zh	33.64	33.79	+0.15	34.66	34.93
en-ja	24.72	25.03	+0.31	34.68	34.89
en-ko	26.15	26.21	+0.06	35.84	35.86
en-ar	39.45	39.22	-0.23	57.60	57.40
en-hi	34.73	34.84	+0.11	54.80	54.88
en-sw	30.77	30.81	+0.04	56.05	56.18
es-en	35.28	35.37	+0.09	60.68	60.72
fr-en	49.86	49.78	-0.08	69.66	69.64
zh-en	32.67	32.57	-0.10	58.16	58.13
ja-en	30.95	30.98	+0.03	56.99	56.95
mean	37.90	37.93	+0.03	56.55	56.58

Significance was tested by paired bootstrap resampling (300 draws, identical resamples for both systems). Across all 16 pairs: zero pairs significantly worse. 68% of outputs are byte-identical to the base model; the remainder are statistically indistinguishable.

Languages

Covers 76 of the 76 languages in Windy Word. Full coverage. Note Tagalog is <2fil>, not <2tl>.

Usage

import ctranslate2
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("WindstormLabs/translate-windy-max")          # tokenizer ships in this repo
tr  = ctranslate2.Translator("WindstormLabs/translate-windy-max", device="cpu", compute_type="int8")

# MADLAD puts the TARGET language in the source as a <2xx> tag.
# Note: Tagalog is <2fil>, not <2tl>.
src = tok.convert_ids_to_tokens(tok.encode("<2es> Where can I find a pharmacy?"))
res = tr.translate_batch([src], beam_size=4)
print(tok.decode(tok.convert_tokens_to_ids(res[0].hypotheses[0]), skip_special_tokens=True))

The tokenizer ships in this repo, so it loads with no network access. (Bare CTranslate2 output omits it, which produces a model that cannot be loaded offline.)

Limitations — stated plainly

• Evaluated on 16 language pairs. Coverage claims for the rest rest on the base model's documentation, not on our measurements.
• FLORES-200 is news and encyclopedic prose. It says little about conversational register, idiom, or domain jargon.
• Quality is inherited from the base model. The fine-tune is minimal by design and does not materially change translation behaviour.
• No human evaluation was performed. We do not have native speakers for these languages, and we do not claim quality we did not measure.

Provenance chain

google/madlad400-3b-mt → CTranslate2 INT8 → LoRA fine-tune (above) → this repo.

Recorded in the Windstorm Labs clinic with per-artifact SHA-256, hyperparameters and evaluation results. Produced on Veron-1 (RTX 5090) on 2026-07-25 by Dr. F.