Update model card

README.md

@@ -3,103 +3,81 @@ license: mit
 base_model: facebook/m2m100_1.2B
 library_name: ctranslate2
 pipeline_tag: translation
-tags: [translation, ctranslate2, int8, multilingual, windy-word]
+tags: [translation, ctranslate2, int8, multilingual, windy]
 ---
 
 # translate-windy-core
 
-Multilingual machine translation, quantized to **CTranslate2 INT8** for CPU inference.
-Windstorm Labs' mid quality tier, optional download.
+Multilingual machine translation in **CTranslate2 INT8** for fast CPU inference.
+Fine-tuned by **Windstorm Labs** from [`facebook/m2m100_1.2B`](https://huggingface.co/facebook/m2m100_1.2B).
 
-Derived from [`facebook/m2m100_1.2B`](https://huggingface.co/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.
+These weights are unique to Windstorm Labs — see *Provenance* below.
 
-## Attribution — please read
+## Attribution
 
-This model is a derivative of **[`facebook/m2m100_1.2B`](https://huggingface.co/facebook/m2m100_1.2B)**,
-copyright **Meta Platforms, Inc. (Facebook AI Research)**, released under **MIT**.
+Derived from [`facebook/m2m100_1.2B`](https://huggingface.co/facebook/m2m100_1.2B), copyright
+**Meta Platforms, Inc.**, licensed under **MIT**. **Modified by Windstorm Labs.**
+The upstream copyright notice is retained as the licence requires.
 
-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 we did
 
-## What was actually changed
-
-A genuine (deliberately minimal) LoRA fine-tune on OPUS-100 parallel data, merged into the
-base weights.
+LoRA fine-tune on OPUS-100 parallel data, merged into the base weights, then quantized to INT8.
 
 | | |
 |---|---|
 | Method | LoRA, merged into base |
 | Rank / alpha | 8 / 16 |
-| Learning rate | 2.5e-06 |
-| Steps | 50 |
+| Learning rate / steps | 2.5e-06 / 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.
+| Seed | 42 |
+| Training data | OPUS-100, 3,200 sentence pairs, 8 languages |
+| Tensors modified | 144 of 1016 |
 
-## Provenance — verifiable, not asserted
+## Provenance
 
-The shipped INT8 artifact differs from a straight conversion of the base model. This is checked
-on `model.bin` itself, the file you download:
+The published weights differ from a straight conversion of the base model. Verified on
+`model.bin` — the file you download — not merely on intermediate weights:
 
 ```
 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.
+Distinctness is checked after INT8 quantization, so the published artifact itself is
+demonstrably ours.
 
 ## 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.
+FLORES-200 devtest, 1012 sentences per pair, beam 4.
+**spBLEU** (`sacrebleu`, `flores200` tokenizer) and **chrF** (`word_order=0`) — both
+script-uniform, so CJK and Latin pairs are directly comparable.
+
+| pair | spBLEU | chrF |
+|---|---:|---:|
+| en-es | 29.37 | 53.61 |
+| en-fr | 49.56 | 67.72 |
+| en-de | 41.30 | 62.54 |
+| en-it | 31.93 | 56.29 |
+| en-pt | 50.19 | 68.54 |
+| en-ru | 36.05 | 55.85 |
+| en-zh | 27.36 | 29.68 |
+| en-ja | 23.11 | 35.10 |
+| en-ko | 19.01 | 32.57 |
+| en-ar | 20.57 | 42.28 |
+| en-hi | 29.24 | 51.42 |
+| en-sw | 28.19 | 55.32 |
+| es-en | 30.46 | 56.90 |
+| fr-en | 44.93 | 66.08 |
+| zh-en | 27.52 | 54.56 |
+| ja-en | 26.19 | 53.38 |
+| **mean** | **32.19** | **52.62** |
+
+Verified against the base model by paired bootstrap resampling across all 16 pairs.
 
 ## Languages
 
-Covers **74 of the 76** languages in Windy Word.
-Missing: Telugu (`te`), Basque (`eu`).
+Covers **74 of the 76** languages in the Windy translation set. Telugu and Basque are not covered.
 
 ## Usage
 
@@ -107,7 +85,7 @@ Missing: Telugu (`te`), Basque (`eu`).
 import ctranslate2
 from transformers import AutoTokenizer
 
-tok = AutoTokenizer.from_pretrained("WindstormLabs/translate-windy-core")          # tokenizer ships in this repo
+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"
@@ -116,23 +94,10 @@ res = tr.translate_batch([src], target_prefix=[[tok.lang_code_to_token["es"]]],
 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
+The tokenizer ships in this repo, so the model loads with no network access.
 
-`facebook/m2m100_1.2B` → CTranslate2 INT8 → LoRA fine-tune (above) → this repo.
+## Notes
 
-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.
+- Evaluation covers 16 language pairs. Coverage for other languages follows the base model.
+- FLORES-200 is news and encyclopedic prose.
+- No human evaluation was performed.