cstr/wmt21-dense-24-wide-x-en-GGUF

WMT21 Dense 24-Wide (X→EN) — GGUF (ggml)

GGUF / ggml conversion of `facebook/wmt21-dense-24-wide-x-en` for use with [CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR).

This is the many-to-English WMT21 competition model — a 4.7B-parameter dense encoder-decoder transformer trained for high-quality translation from 7 source languages into English. It won the WMT21 shared task for multiple language pairs. Architecture: 24 encoder + 24 decoder layers (d=2048, 32 heads, FFN=16384, ReLU, pre-norm, sinusoidal positions). Distributed under MIT license.

Files

Supported languages (7 → English)

Target is always English (en).

Quick start

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF
cmake --build build -j

# 2. Pull model
huggingface-cli download cstr/wmt21-dense-24-wide-x-en-GGUF wmt21-dense-24-wide-x-en-q8_0.gguf --local-dir .

# 3. Translate
./build/bin/crispasr --backend m2m100 -m wmt21-dense-24-wide-x-en-q8_0.gguf \
    --text "Die Maschine lernt schnell und verändert die Welt." \
    -sl de -tl en

# Japanese → English
./build/bin/crispasr --backend m2m100 -m wmt21-dense-24-wide-x-en-q8_0.gguf \
    --text "機械学習は世界を変えています。" \
    -sl ja -tl en

Quality verification

Architecture

Text → SentencePiece BPE tokenizer (128K vocab, 8 lang codes)
     → Source lang token (__de__) + text tokens + </s>
     → 24-layer transformer encoder (d=2048, 32 heads, FFN=16384, ReLU, pre-norm)
     → Sinusoidal positional embeddings (pre-computed)
     → 24-layer transformer decoder (self-attn + cross-attn + FFN)
     → Shared embedding LM head (tied weights)
     → English forced as first decoder token
     → Greedy decode → translated English text

Conversion

python models/convert-m2m100-to-gguf.py \
    --input facebook/wmt21-dense-24-wide-x-en \
    --output wmt21-dense-24-wide-x-en-f16.gguf

Related models

• —`cstr/wmt21-dense-24-wide-en-x-GGUF` — English-to-many (reverse direction)
• —`cstr/m2m100-418m-GGUF` — smaller 100-language any-to-any model
• —`facebook/wmt21-dense-24-wide-x-en` — original PyTorch model

Provenance and EU AI Act Art. 53 note

• —Upstream model: facebook/wmt21-dense-24-wide-x-en — published by facebook.
• —Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• —What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• —Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• —Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.