Afri model addition

README.md

@@ -68,4 +68,20 @@ docker build -t translator-api .
 
 # Run Docker container mapping port 7860
 docker run -p 7860:7860 -v $(pwd)/data:/data translator-api
-```
+```
+
+## 📜 Citation & Credits
+
+This Space utilizes [AfriNLLB](https://huggingface.co/AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd) developed by AfriNLP:
+
+```bibtex
+@inproceedings{moslem-etal-2026-afrinllb,
+    title = "{A}fri{NLLB}: Efficient Translation Models for African Languages",
+    author = "Moslem, Yasmin and Wassie, Aman Kassahun and Gizachew, Amanuel",
+    booktitle = "Proceedings of the Seventh Workshop on African Natural Language Processing (AfricaNLP)",
+    month = mar,
+    year = "2026",
+    address = "Rabat, Morocco",
+    publisher = "Association for Computational Linguistics",
+    url = "https://openreview.net/forum?id=hVJZNUZBur"
+}

backend/providers/__init__.py

@@ -13,8 +13,9 @@ from providers.base import ProviderInfo, TranslationProvider
 from providers.gemini import GeminiProvider
 from providers.groq import GroqProvider, GroqQwenProvider
 from providers.madlad import MADLADProvider
-from providers.nllb import NLLBProvider, SunbirdNLLBProvider
+from providers.nllb import NLLBProvider
 from providers.ollama import OllamaProvider
+from providers.gemma import SwahiliGemmaProvider
 
 _PROVIDERS: list[TranslationProvider] = [
     NLLBProvider(),
@@ -26,8 +27,18 @@ _PROVIDERS: list[TranslationProvider] = [
         ),
         hf_model=os.environ.get("HF_MODEL_1_3B", "facebook/nllb-200-distilled-1.3B"),
     ),
-    SunbirdNLLBProvider(),
+    NLLBProvider(
+        provider_id="afrinllb",
+        name="AfriNLLB (600M)",
+        model_dir=os.environ.get(
+            "CT2_MODEL_DIR_AFRINLLB", "/models/afrinllb-600m-int8"
+        ),
+        hf_model=os.environ.get(
+            "HF_MODEL_AFRINLLB", "AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd"
+        ),
+    ),
     MADLADProvider(),
+    SwahiliGemmaProvider(),
     OllamaProvider(),
     GeminiProvider(),
     GroqQwenProvider(),

backend/providers/gemma.py

@@ -0,0 +1,65 @@
+"""Swahili Gemma 1B engine via llama-cpp-python (GGUF)."""
+
+from __future__ import annotations
+
+import os
+from threading import Lock
+
+from providers.base import TranslationProvider
+
+MODEL_PATH = os.environ.get(
+    "GEMMA_MODEL_PATH",
+    "/data/models/swahili-gemma-1b/Q4_K_M/swahili-gemma-1b-q4_k_m.gguf",
+)
+
+
+class SwahiliGemmaProvider(TranslationProvider):
+    kind = "local"
+    description = "Fine-tuned Gemma 1B · Specialized English -> Swahili translation"
+    private = True
+    setup_hint = "Download GGUF model to /data/models/swahili-gemma-1b"
+
+    def __init__(
+        self,
+        provider_id: str = "swahili_gemma",
+        name: str = "Swahili Gemma (1B GGUF)",
+        model_path: str = MODEL_PATH,
+    ) -> None:
+        self.id = provider_id
+        self.name = name
+        self.model_path = model_path
+        self._llm = None
+        self._lock = Lock()
+
+    def is_available(self) -> bool:
+        return os.path.isfile(self.model_path)
+
+    def _ensure_loaded(self) -> None:
+        if self._llm is not None:
+            return
+        with self._lock:
+            if self._llm is not None:
+                return
+            from llama_cpp import Llama
+
+            self._llm = Llama(
+                model_path=self.model_path,
+                n_ctx=2048,
+                n_threads=8,  # Match your 8 vCPUs
+                verbose=False,
+            )
+
+    def translate(self, text: str, src: str, tgt: str, api_key: str | None = None) -> str:
+        self._ensure_loaded()
+        
+        prompt = f"Translate to Swahili: {text}"
+        
+        response = self._llm(
+            prompt,
+            max_tokens=256,
+            temperature=0.3,
+            top_p=0.95,
+            repeat_penalty=1.1,
+        )
+        
+        return response["choices"][0]["text"].strip()

backend/requirements.txt

@@ -9,7 +9,9 @@ ctranslate2==4.5.0
 transformers==4.47.1
 sentencepiece==0.2.0
 protobuf==5.29.2
+llama-cpp-python>=0.2.80
 
 # ctranslate2 4.5 imports pkg_resources, removed in setuptools>=81
 setuptools<81
 scalar-fastapi
+huggingface_hub

entrypoint.sh

@@ -15,10 +15,13 @@ HF_MODEL="google/madlad400-3b-mt" \
 CT2_MODEL_DIR="/data/models/madlad400-3b-mt-int8" \
 python backend/convert_model.py || echo "WARNING: MADLAD conversion failed, continuing"
 
-echo "== Converting Sunbird NLLB 3.3B =="
-HF_MODEL="Sunbird/translate-nllb-3.3b-salt" \
-CT2_MODEL_DIR="/data/models/sunbird-nllb-3.3b-int8" \
-python backend/convert_model.py || echo "WARNING: Sunbird NLLB 3.3B conversion failed, continuing"
+echo "== Converting AfriNLLB 600M =="
+HF_MODEL="AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd" \
+CT2_MODEL_DIR="/data/models/afrinllb-600m-int8" \
+python backend/convert_model.py || echo "WARNING: AfriNLLB conversion failed, continuing"
+
+echo "== Downloading Swahili Gemma GGUF =="
+python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='CraneAILabs/swahili-gemma-1b-GGUF', filename='Q4_K_M/swahili-gemma-1b-q4_k_m.gguf', local_dir='/data/models/swahili-gemma-1b')" || echo "WARNING: Swahili Gemma download failed, continuing"
 
 echo "== Starting server =="
 cd backend && exec uvicorn main:app --host 0.0.0.0 --port 7860