HUGGING_FACE_SPACE_TRAINING.md

Hugging Face Space Training

This project keeps internal orchestration as an external orchestration microservice. Model training is handled by a separate Hugging Face Docker Space:

mortadhabbb/train_model_chatbot

The local Django chatbot generates service_intents_model_training.csv, uploads it to the Space, and the Space trains a service-intent extractor with Hugging Face resources.

Space Code

The local clone prepared for upload is:

D:\Nouveau dossier\huggingface_spaces\train_model_chatbot

It contains:

• app.py: FastAPI job API.
• training/train_service_intent_extractor.py: standalone trainer.
• Dockerfile: Hugging Face Docker Space runtime.
• requirements.txt: Python dependencies.

Space Secrets

Set these in the Hugging Face Space settings when needed:

TRAINING_API_KEY=change-me
HF_TOKEN=hf_...

TRAINING_API_KEY protects the training API. HF_TOKEN is only required if the Space should push the trained model to a Hugging Face model repository.

Local Settings

Configure these in .env:

SERVICE_INTENT_TRAINING_SPACE_URL=https://mortadhabbb-train-model-chatbot.hf.space
SERVICE_INTENT_TRAINING_SPACE_API_KEY=
SERVICE_INTENT_TRAINING_OUTPUT_REPO_ID=
SERVICE_INTENT_TRAINING_POLL_SECONDS=20

Submit Training From Django

Rebuild the datasets and submit a remote training job:

python manage.py submit_remote_service_intent_training --rebuild-datasets

Submit and wait until the Space finishes:

python manage.py submit_remote_service_intent_training --rebuild-datasets --wait

Wait and download the trained artifact ZIP:

python manage.py submit_remote_service_intent_training --rebuild-datasets --wait --download-artifact artifacts/service_intent_model_remote.zip

Ask the Space to push the trained model to a Hugging Face model repo:

python manage.py submit_remote_service_intent_training --rebuild-datasets --wait --push-to-hub --output-repo-id mortadhabbb/service-intent-extractor

For free CPU Spaces, start with a smaller model:

python manage.py submit_remote_service_intent_training --model-id google/flan-t5-small --rebuild-datasets --wait

Use google/flan-t5-base when the Space has enough CPU/GPU memory and time.