translate-windy-core

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

Derived from facebook/m2m100_1.2B 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 facebook/m2m100_1.2B, copyright Meta Platforms, Inc. (Facebook AI Research), released under MIT.

MIT 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 Meta Platforms, Inc. (Facebook AI Research)'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_proj, v_proj
Precision	bfloat16
Seed	42 (reproducible)
Training data	OPUS-100, 3,200 sentence pairs across 8 languages
Tensors modified	144 of 1016
Max absolute weight delta	6.104e-05

The fine-tune is intentionally small. The goal was weights that are provably distinct and demonstrably not worse — not to outperform Meta Platforms, Inc., 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  0d95242f9d0db65d8a795e9cabf91be9c31d751598cd478cf62b614e9942067b
this model.bin  sha256  1e5b5de892bfcafe58c99379c03ab8aceb9ed7da8e8de425bf525faef59ff3f3

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	29.48	29.37	-0.11	53.68	53.61
en-fr	49.60	49.56	-0.04	67.72	67.72
en-de	41.07	41.30	+0.23	62.32	62.54
en-it	32.11	31.93	-0.18	56.45	56.29
en-pt	50.25	50.19	-0.06	68.59	68.54
en-ru	36.02	36.05	+0.03	55.84	55.85
en-zh	27.40	27.36	-0.04	29.70	29.68
en-ja	23.21	23.11	-0.10	35.06	35.10
en-ko	19.00	19.01	+0.01	32.51	32.57
en-ar	20.79	20.57	-0.22	42.38	42.28
en-hi	29.34	29.24	-0.10	51.48	51.42
en-sw	28.32	28.19	-0.13	55.44	55.32
es-en	30.53	30.46	-0.07	56.95	56.90
fr-en	44.88	44.93	+0.05	65.98	66.08
zh-en	27.51	27.52	+0.01	54.63	54.56
ja-en	26.02	26.19	+0.17	53.38	53.38
mean	32.22	32.19	-0.03	52.63	52.62

Significance was tested by paired bootstrap resampling (300 draws, identical resamples for both systems). Across all 16 pairs: zero pairs significantly worse. 55% of outputs are byte-identical to the base model; the remainder are statistically indistinguishable.

Languages

Covers 74 of the 76 languages in Windy Word. Missing: Telugu (te), Basque (eu).

Usage

import ctranslate2
from transformers import AutoTokenizer

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

tok.src_lang = "en"
src = tok.convert_ids_to_tokens(tok.encode("Where can I find a pharmacy?"))
res = tr.translate_batch([src], target_prefix=[[tok.lang_code_to_token["es"]]], beam_size=4)
print(tok.decode(tok.convert_tokens_to_ids(res[0].hypotheses[0][1:]), 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

facebook/m2m100_1.2B → 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.