MADLAD-400 10B MT — GGUF (ggml)

GGUF / ggml conversion of google/madlad400-10b-mt for use with CrispStrobe/CrispASR.

MADLAD-400 10B MT is the largest and most capable model in Google's MADLAD-400 family. With 10.7B parameters across 32 wide layers (hidden dimension 4096), it delivers top-tier translation accuracy across 450+ languages, trained on 1 trillion tokens. Distributed under the Apache 2.0 license.

Files Available

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/madlad400-10b-mt-GGUF madlad400-10b-mt-q4_k.gguf --local-dir .

# 3. Translate (uses <2xx> language tags)
./build/bin/crispasr --backend madlad -m madlad400-10b-mt-q4_k.gguf \
    --text "Hello world, how are you today?" \
    -sl en -tl de

# English → Japanese
./build/bin/crispasr --backend madlad -m madlad400-10b-mt-q4_k.gguf \
    --text "Machine learning is changing the world." \
    -sl en -tl ja

# French → Portuguese
./build/bin/crispasr --backend madlad -m madlad400-10b-mt-q4_k.gguf \
    --text "Bonjour le monde!" \
    -sl fr -tl pt

Supported languages (450+)

MADLAD-400 supports over 450 languages using <2xx> target language tags (ISO 639 codes). This includes all major world languages plus many low-resource languages. See the original model card for the full language list.

Architecture

Text → SentencePiece tokenizer (256K vocab, shared encoder-decoder)
     → <2xx> target language tag prepended to source text
     → T5 encoder (32 layers, d=4096, 10.7B parameters)
     → T5 decoder (32 layers, d=4096) with cross-attention
     → Greedy decode → translated text

Parity Verification

Converted and verified with CrispASR per-stage parity test against PyTorch reference outputs:

crispasr-diff madlad madlad400-10b-mt-q4_k.gguf --ref madlad400-10b-mt-ref.gguf

Related models

• cstr/madlad400-3b-mt-GGUF — MADLAD-400 3B MT
• cstr/madlad400-7b-mt-GGUF — MADLAD-400 7B MT (standard)
• cstr/madlad400-7b-mt-bt-GGUF — MADLAD-400 7B MT (Back-Translation fine-tuned)
• cstr/m2m100-418m-GGUF — M2M-100, 100 languages, any-to-any
• google/madlad400-10b-mt — original PyTorch model

Citation

@misc{kudugunta2023madlad400,
      title={MADLAD-400: A Multilingual And Document-Level Large Audited Dataset}, 
      author={Sneha Kudugunta and Isaac Caswell and Biao Zhang and Xavier Garcia and Christopher A. Choquette-Choo and Katherine Lee and Derrick Xin and Aditya Kusupati and Romi Stella and Ankur Bapna and Orhan Firat},
      year={2023},
      eprint={2309.04662},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}