French → Wolof Marian / Candle (Malinali)
Public on-device pack for Malinali. Fine-tuned MarianMT
weights in safetensors form plus fast tokenizers for Candle (marian_flutter).
Upstream credit
- Fine-tune:
makhtar7186/traduction_fr_wolof - Base model:
Helsinki-NLP/opus-mt-fr-en - Dataset: galsenai/french-wolof-translation
- Base project: Helsinki-NLP / OPUS-MT
The upstream model card does not declare a license. OPUS-MT base weights are typically CC-BY 4.0. This repo only repackages the root weights and converts SentencePiece to Hugging Face fast tokenizer JSON. Malinali does not claim ownership of the trained model.
Tokenizer
source.spm, target.spm, and vocab.json are the opus-mt-fr-en pair
(vocab size 59514, pad id 59513). The fine-tune did not add Wolof pieces.
tokenizer_config.json still says target_lang: en; there is no >>en<<
or >>wo<< token, so no language code is prefixed. Do not resize the vocab.
Files (required)
| File | Role |
|---|---|
config.json | Marian config |
model.safetensors | Weights (root file only; not training checkpoints) |
tokenizer-enc.json | French (source) fast tokenizer |
tokenizer-dec.json | Target fast tokenizer (same inventory as the source SPM vocab) |
Direction: French → Wolof.
Quality
Best eval_bleu logged by the upstream Trainer: 9.32 at epoch 18
(checkpoint-8010), on the fine-tune held-out split (about 1,800 sentences
from galsenai/french-wolof-translation). Epoch 20 scored 9.28 and early
stopping ended the run. Malinali shows this as BLEU 9.3 / 100.
This is the training run's own metric, not a new SacreBLEU computed by Malinali.