Upload all 119 int8 CTranslate2 pairs

.gitattributes

@@ -33,3 +33,241 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+aghem-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+aghem-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ajamiya-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ajamiya-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+akoose-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+akoose-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+awing-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+awing-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+babanki-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+babanki-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+bafia-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+bafia-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+bakoko-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+bakoko-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+bakweri-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+bakweri-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+bidwee-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+bidwee-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+bulu-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+bulu-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+bum-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+bum-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+cuvok-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+cuvok-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+denya-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+denya-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+dii-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+dii-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+doyayo-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+doyayo-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ejagham-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ejagham-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+english-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+english-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+esimbi-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+esimbi-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ewondo-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ewondo-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-aghem/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-aghem/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ajamiya/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ajamiya/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-awing/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-awing/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-babanki/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-babanki/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-bafia/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-bafia/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-bakoko/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-bakoko/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-bakweri/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-bakweri/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-bidwee/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-bidwee/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-bulu/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-bulu/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-bum/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-bum/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-cuvok/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-cuvok/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-denya/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-denya/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-dii/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-dii/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-doyayo/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-doyayo/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ejagham/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ejagham/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-english/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-english/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-esimbi/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-esimbi/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ewondo/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ewondo/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-fufulde/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-fufulde/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-gbaya/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-gbaya/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ghomala/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ghomala/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-guidar/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-guidar/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-guiziga/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-guiziga/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-isu/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-isu/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-kapsiki/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-kapsiki/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-kenyang/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-kenyang/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-koonzime/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-koonzime/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-lamnso/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-lamnso/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-limbum/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-limbum/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mankon/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mankon/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-massana/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-massana/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mbembe/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mbembe/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-medumba/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-medumba/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-meta/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-meta/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mmen/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mmen/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mofa/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mofa/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mofu/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mofu/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-moghamo/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-moghamo/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mpumpong/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mpumpong/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-mundani/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-mundani/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngi/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngi/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngienboum/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngienboum/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngomba/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngomba/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngombale/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngombale/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngwo/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-ngwo/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-nomaande/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-nomaande/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-nugunu/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-nugunu/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-oku/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-oku/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-pana/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-pana/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-peere/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-peere/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-pinyin/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-pinyin/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-punu/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-punu/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-samba/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-samba/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-tunen/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-tunen/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-tupuri/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-tupuri/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-vute/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-vute/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-weh/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-weh/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-yambeta/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-yambeta/target.spm filter=lfs diff=lfs merge=lfs -text
+francais-yemba/source.spm filter=lfs diff=lfs merge=lfs -text
+francais-yemba/target.spm filter=lfs diff=lfs merge=lfs -text
+fufulde-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+fufulde-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+gbaya-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+gbaya-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ghomala-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ghomala-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+guidar-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+guidar-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+guiziga-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+guiziga-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+isu-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+isu-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+kapsiki-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+kapsiki-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+kenyang-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+kenyang-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+koonzime-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+koonzime-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+lamnso-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+lamnso-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+limbum-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+limbum-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mankon-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mankon-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+massana-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+massana-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mbembe-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mbembe-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+medumba-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+medumba-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+meta-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+meta-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mmen-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mmen-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mofa-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mofa-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mofu-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mofu-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+moghamo-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+moghamo-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mpumpong-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mpumpong-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+mundani-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+mundani-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ngi-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ngi-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ngienboum-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ngienboum-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ngomba-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ngomba-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ngombale-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ngombale-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+ngwo-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+ngwo-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+nomaande-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+nomaande-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+nugunu-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+nugunu-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+oku-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+oku-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+pana-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+pana-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+peere-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+peere-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+pinyin-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+pinyin-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+punu-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+punu-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+samba-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+samba-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+tunen-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+tunen-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+tupuri-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+tupuri-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+vute-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+vute-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+weh-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+weh-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+yambeta-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+yambeta-francais/target.spm filter=lfs diff=lfs merge=lfs -text
+yemba-francais/source.spm filter=lfs diff=lfs merge=lfs -text
+yemba-francais/target.spm filter=lfs diff=lfs merge=lfs -text

aghem-francais/README.md

@@ -0,0 +1,41 @@
+---
+license: cc-by-4.0
+library_name: ctranslate2
+tags:
+  - translation
+  - ctranslate2
+  - int8
+  - marian
+  - cameroon
+  - low-resource
+language:
+  - fr
+---
+
+# aghem-francais — int8 (CTranslate2)
+
+Compressed French-pivot translation model for Cameroonian languages, part of the
+[Lingo / NativeAI](https://lingo.cm) language-preservation project.
+
+This is an **int8 [CTranslate2](https://github.com/OpenNMT/CTranslate2) conversion**
+of our fp32 MarianMT model — about **3.8× smaller** (~285 MB → ~75 MB) and
+**~6× faster on CPU**, so dozens of languages can be served on free-tier
+hardware. The quality difference from fp32 is below the noise floor of these
+narrow-domain, low-resource models (see the
+[research log](https://lingo.cm/blog)). The fp32 originals are archived.
+
+## Usage
+
+```python
+import ctranslate2
+from transformers import MarianTokenizer
+
+translator = ctranslate2.Translator("aghem-francais", device="cpu", compute_type="int8")
+tok = MarianTokenizer.from_pretrained("aghem-francais")
+
+src = tok.convert_ids_to_tokens(tok.encode("Bonjour, comment vas-tu ?"))
+out = translator.translate_batch([src], beam_size=4)
+print(tok.decode(tok.convert_tokens_to_ids(out[0].hypotheses[0]), skip_special_tokens=True))
+```
+
+Translations chain through French (the pivot): e.g. Ghomálá' → français → Ewondo.

aghem-francais/config.json

@@ -0,0 +1,10 @@
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

aghem-francais/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cca9a2d6feef351a69f011fe7cde3fbcf1a59ac86b3dd51620a783a7f5344bf9
+size 76714395

aghem-francais/source.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:173e9f493a668fe396d599e28d414a201193094e6ffd7a4678e5aab0f6d3d838
+size 778395

aghem-francais/special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

aghem-francais/target.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d0e717c77053f1c4b856d8661d9cb87c64f083a35418c087b9146300e4f585
+size 802397

aghem-francais/tokenizer_config.json

@@ -0,0 +1,38 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59513": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "</s>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": "en",
+  "sp_model_kwargs": {},
+  "target_lang": "fr",
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}

ajamiya-francais/README.md

@@ -0,0 +1,41 @@
+---
+license: cc-by-4.0
+library_name: ctranslate2
+tags:
+  - translation
+  - ctranslate2
+  - int8
+  - marian
+  - cameroon
+  - low-resource
+language:
+  - fr
+---
+
+# ajamiya-francais — int8 (CTranslate2)
+
+Compressed French-pivot translation model for Cameroonian languages, part of the
+[Lingo / NativeAI](https://lingo.cm) language-preservation project.
+
+This is an **int8 [CTranslate2](https://github.com/OpenNMT/CTranslate2) conversion**
+of our fp32 MarianMT model — about **3.8× smaller** (~285 MB → ~75 MB) and
+**~6× faster on CPU**, so dozens of languages can be served on free-tier
+hardware. The quality difference from fp32 is below the noise floor of these
+narrow-domain, low-resource models (see the
+[research log](https://lingo.cm/blog)). The fp32 originals are archived.
+
+## Usage
+
+```python
+import ctranslate2
+from transformers import MarianTokenizer
+
+translator = ctranslate2.Translator("ajamiya-francais", device="cpu", compute_type="int8")
+tok = MarianTokenizer.from_pretrained("ajamiya-francais")
+
+src = tok.convert_ids_to_tokens(tok.encode("Bonjour, comment vas-tu ?"))
+out = translator.translate_batch([src], beam_size=4)
+print(tok.decode(tok.convert_tokens_to_ids(out[0].hypotheses[0]), skip_special_tokens=True))
+```
+
+Translations chain through French (the pivot): e.g. Ghomálá' → français → Ewondo.

ajamiya-francais/config.json

@@ -0,0 +1,10 @@
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

ajamiya-francais/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdcc3ef0e3a4fe8f6376917faa0d31bd0a1677fe528138b0963d00bb3121e1fd
+size 76714395

ajamiya-francais/source.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:173e9f493a668fe396d599e28d414a201193094e6ffd7a4678e5aab0f6d3d838
+size 778395

ajamiya-francais/special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

ajamiya-francais/target.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d0e717c77053f1c4b856d8661d9cb87c64f083a35418c087b9146300e4f585
+size 802397

ajamiya-francais/tokenizer_config.json

@@ -0,0 +1,38 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59513": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "</s>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": "en",
+  "sp_model_kwargs": {},
+  "target_lang": "fr",
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}

akoose-francais/README.md

@@ -0,0 +1,41 @@
+---
+license: cc-by-4.0
+library_name: ctranslate2
+tags:
+  - translation
+  - ctranslate2
+  - int8
+  - marian
+  - cameroon
+  - low-resource
+language:
+  - fr
+---
+
+# akoose-francais — int8 (CTranslate2)
+
+Compressed French-pivot translation model for Cameroonian languages, part of the
+[Lingo / NativeAI](https://lingo.cm) language-preservation project.
+
+This is an **int8 [CTranslate2](https://github.com/OpenNMT/CTranslate2) conversion**
+of our fp32 MarianMT model — about **3.8× smaller** (~285 MB → ~75 MB) and
+**~6× faster on CPU**, so dozens of languages can be served on free-tier
+hardware. The quality difference from fp32 is below the noise floor of these
+narrow-domain, low-resource models (see the
+[research log](https://lingo.cm/blog)). The fp32 originals are archived.
+
+## Usage
+
+```python
+import ctranslate2
+from transformers import MarianTokenizer
+
+translator = ctranslate2.Translator("akoose-francais", device="cpu", compute_type="int8")
+tok = MarianTokenizer.from_pretrained("akoose-francais")
+
+src = tok.convert_ids_to_tokens(tok.encode("Bonjour, comment vas-tu ?"))
+out = translator.translate_batch([src], beam_size=4)
+print(tok.decode(tok.convert_tokens_to_ids(out[0].hypotheses[0]), skip_special_tokens=True))
+```
+
+Translations chain through French (the pivot): e.g. Ghomálá' → français → Ewondo.

akoose-francais/config.json

@@ -0,0 +1,10 @@
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

akoose-francais/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc5d95406f17cec316288433d7a763894606036288c93fb722f9cd1feab8d0bc
+size 76714395

akoose-francais/source.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:173e9f493a668fe396d599e28d414a201193094e6ffd7a4678e5aab0f6d3d838
+size 778395

akoose-francais/special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

akoose-francais/target.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d0e717c77053f1c4b856d8661d9cb87c64f083a35418c087b9146300e4f585
+size 802397

akoose-francais/tokenizer_config.json

@@ -0,0 +1,38 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59513": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "</s>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": "en",
+  "sp_model_kwargs": {},
+  "target_lang": "fr",
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}

awing-francais/README.md

@@ -0,0 +1,41 @@
+---
+license: cc-by-4.0
+library_name: ctranslate2
+tags:
+  - translation
+  - ctranslate2
+  - int8
+  - marian
+  - cameroon
+  - low-resource
+language:
+  - fr
+---
+
+# awing-francais — int8 (CTranslate2)
+
+Compressed French-pivot translation model for Cameroonian languages, part of the
+[Lingo / NativeAI](https://lingo.cm) language-preservation project.
+
+This is an **int8 [CTranslate2](https://github.com/OpenNMT/CTranslate2) conversion**
+of our fp32 MarianMT model — about **3.8× smaller** (~285 MB → ~75 MB) and
+**~6× faster on CPU**, so dozens of languages can be served on free-tier
+hardware. The quality difference from fp32 is below the noise floor of these
+narrow-domain, low-resource models (see the
+[research log](https://lingo.cm/blog)). The fp32 originals are archived.
+
+## Usage
+
+```python
+import ctranslate2
+from transformers import MarianTokenizer
+
+translator = ctranslate2.Translator("awing-francais", device="cpu", compute_type="int8")
+tok = MarianTokenizer.from_pretrained("awing-francais")
+
+src = tok.convert_ids_to_tokens(tok.encode("Bonjour, comment vas-tu ?"))
+out = translator.translate_batch([src], beam_size=4)
+print(tok.decode(tok.convert_tokens_to_ids(out[0].hypotheses[0]), skip_special_tokens=True))
+```
+
+Translations chain through French (the pivot): e.g. Ghomálá' → français → Ewondo.

awing-francais/config.json

@@ -0,0 +1,10 @@
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

awing-francais/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78672e6fe316189e10fc9ed1c694920003f9c764a54ea50713604e90eed0e0f2
+size 76714395

awing-francais/source.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:173e9f493a668fe396d599e28d414a201193094e6ffd7a4678e5aab0f6d3d838
+size 778395

awing-francais/special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

awing-francais/target.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d0e717c77053f1c4b856d8661d9cb87c64f083a35418c087b9146300e4f585
+size 802397

awing-francais/tokenizer_config.json

@@ -0,0 +1,38 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59513": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "</s>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": "en",
+  "sp_model_kwargs": {},
+  "target_lang": "fr",
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}

babanki-francais/README.md

@@ -0,0 +1,41 @@
+---
+license: cc-by-4.0
+library_name: ctranslate2
+tags:
+  - translation
+  - ctranslate2
+  - int8
+  - marian
+  - cameroon
+  - low-resource
+language:
+  - fr
+---
+
+# babanki-francais — int8 (CTranslate2)
+
+Compressed French-pivot translation model for Cameroonian languages, part of the
+[Lingo / NativeAI](https://lingo.cm) language-preservation project.
+
+This is an **int8 [CTranslate2](https://github.com/OpenNMT/CTranslate2) conversion**
+of our fp32 MarianMT model — about **3.8× smaller** (~285 MB → ~75 MB) and
+**~6× faster on CPU**, so dozens of languages can be served on free-tier
+hardware. The quality difference from fp32 is below the noise floor of these
+narrow-domain, low-resource models (see the
+[research log](https://lingo.cm/blog)). The fp32 originals are archived.
+
+## Usage
+
+```python
+import ctranslate2
+from transformers import MarianTokenizer
+
+translator = ctranslate2.Translator("babanki-francais", device="cpu", compute_type="int8")
+tok = MarianTokenizer.from_pretrained("babanki-francais")
+
+src = tok.convert_ids_to_tokens(tok.encode("Bonjour, comment vas-tu ?"))
+out = translator.translate_batch([src], beam_size=4)
+print(tok.decode(tok.convert_tokens_to_ids(out[0].hypotheses[0]), skip_special_tokens=True))
+```
+
+Translations chain through French (the pivot): e.g. Ghomálá' → français → Ewondo.

babanki-francais/config.json

@@ -0,0 +1,10 @@
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

babanki-francais/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3325a5fc9d0277cdbd92c6b012597a014db604d215853c712753ac89f111d96
+size 76714395

babanki-francais/source.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:173e9f493a668fe396d599e28d414a201193094e6ffd7a4678e5aab0f6d3d838
+size 778395

babanki-francais/special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

babanki-francais/target.spm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d0e717c77053f1c4b856d8661d9cb87c64f083a35418c087b9146300e4f585
+size 802397

babanki-francais/tokenizer_config.json

@@ -0,0 +1,38 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "59513": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "</s>",
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": "en",
+  "sp_model_kwargs": {},
+  "target_lang": "fr",
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}

bafia-francais/README.md

@@ -0,0 +1,41 @@
+---
+license: cc-by-4.0
+library_name: ctranslate2
+tags:
+  - translation
+  - ctranslate2
+  - int8
+  - marian
+  - cameroon
+  - low-resource
+language:
+  - fr
+---
+
+# bafia-francais — int8 (CTranslate2)
+
+Compressed French-pivot translation model for Cameroonian languages, part of the
+[Lingo / NativeAI](https://lingo.cm) language-preservation project.
+
+This is an **int8 [CTranslate2](https://github.com/OpenNMT/CTranslate2) conversion**
+of our fp32 MarianMT model — about **3.8× smaller** (~285 MB → ~75 MB) and
+**~6× faster on CPU**, so dozens of languages can be served on free-tier
+hardware. The quality difference from fp32 is below the noise floor of these
+narrow-domain, low-resource models (see the
+[research log](https://lingo.cm/blog)). The fp32 originals are archived.
+
+## Usage
+
+```python
+import ctranslate2
+from transformers import MarianTokenizer
+
+translator = ctranslate2.Translator("bafia-francais", device="cpu", compute_type="int8")
+tok = MarianTokenizer.from_pretrained("bafia-francais")
+
+src = tok.convert_ids_to_tokens(tok.encode("Bonjour, comment vas-tu ?"))
+out = translator.translate_batch([src], beam_size=4)
+print(tok.decode(tok.convert_tokens_to_ids(out[0].hypotheses[0]), skip_special_tokens=True))
+```
+
+Translations chain through French (the pivot): e.g. Ghomálá' → français → Ewondo.

bafia-francais/config.json

@@ -0,0 +1,10 @@
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

bafia-francais/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f07ec38595fe9cc9158100b9af30a8f5a1149335877aca36bf185364b58c0781
+size 76714395