Model Description

This machine translation model translates text from Dyula to French. It is built on a fairseq model architecture proposed by Facebook. The architecture was replicated using a Dyula-French translation dataset created by data354. The model was later quantized into int8 and exported to ctranslate2 format for fast inference. The model is designed to support a variety of educational applications by providing accurate and contextually relevant translations between these languages.

Intended Use

The model is specifically designed to support AI Student Learning Assistant (AISLA), a free educational tool aimed at helping students learn and communicate in their native language.

The model is particularly valuable for enhancing educational accessibility for Dyula-speaking students by enabling reliable translations from Dyula to French. It is intended to be integrated into platforms like Discord to provide seamless support within educational environments.

Deployment

This folder contains the resources required for deploying the trained model onto Highwind.

Usage

All commands below are run from the deployment directory.

Building the Model Image

This step builds the Kserve predictor image that contains the model.

1. Ensure the trained model folder model_dir contains:

   • model.bin (the model itself)
   • config.json (model configuration file)
   • combined_model_2000.model (sentence piece model for tokenization)
   • shared_vocabulary.json (shared vocabulary)

2. The deployment folder should include:

   • Dockerfile (steps to build the image)
   • main.py (starts the Kserve server and runs the model for translation inference tasks)
   • serve-requirements.txt (model dependencies)

3. Build the container locally without caching and tag it:

   docker build --no-cache -t dyula-french-seqf_8-25-5beams:latest .

Local Testing

1. After building the Kserve predictor image, spin it up to test the model inference:

   docker compose up -d
   docker compose logs

2. Send a payload to the model to test its response using the curl command to send a POST request with an example JSON payload.

   Run this from another terminal (remember to navigate to this folder first)

   Linux/Mac Bash/zsh:

   curl -X POST http://localhost:8080/v2/models/model/infer -H 'Content-Type: application/json' -d @./input.json

   Windows PowerShell:

   $json = Get-Content -Raw -Path ./input.json
   $response = Invoke-WebRequest -Uri http://localhost:8080/v2/models/model/infer -Method Post -ContentType 'application/json' -Body ([System.Text.Encoding]::UTF8.GetBytes($json))
   $responseObject = $response.Content | ConvertFrom-Json
   $responseObject | ConvertTo-Json -Depth 10

   Using Postman: Send a POST request to http://localhost:8080/v2/models/model/infer with a JSON body:

   {
       "inputs": [
           {
               "name": "input-0",
               "shape": [1],
               "datatype": "BYTES",
               "parameters": null,
               "data": ["i tɔgɔ bi cogodɔ"]
           }
       ]
   }

Deployment on Highwind

Once the translation is successfully generated through a POST request, proceed to deploy the model on Highwind:

1. Create a new asset in Highwind and follow the push commands.
2. Create a new use case and add the asset to it.
3. Deploy the use case.

Now, the model is ready for inference through API calls!