QUICKSTART.md

Quick Start - Docker Space Deployment

Files to Upload

Required files for HF Space:

Dockerfile
train_nllb_hf_spaces.py
requirements_training.txt
run_training.sh

3-Step Setup

1. Create Space

• URL: https://huggingface.co/spaces
• SDK: Docker (not Gradio)
• Hardware: T4 medium

2. Configure

Edit run_training.sh:

HF_USERNAME="your-username"          # Your HF username
LANGUAGE_PAIRS="en-fr en-de en-es"   # Language pairs
MAX_SAMPLES="50000"                  # Samples per pair
EPOCHS="3"                           # Training epochs

Set secret in Space Settings → Repository Secrets:

HF_TOKEN = your-token-from-hf-settings

3. Deploy

git clone https://huggingface.co/spaces/YOUR_USERNAME/your-space-name
cd your-space-name
cp /path/to/files/* .
git add .
git commit -m "Initial setup"
git push

Monitoring

• Logs: Space → Logs tab (real-time)
• Checkpoints: https://huggingface.co/YOUR_USERNAME/fine-tuned-nllb-600M
• First checkpoint: ~10-15 minutes

Time Estimates

Config	Time
2 pairs × 20k (test)	45 min
3 pairs × 50k	2-3 hrs
5 pairs × 100k	4-6 hrs

Resume After Crash

Edit run_training.sh:

RESUME_FROM="your-username/fine-tuned-nllb-600M"

Push changes. Training resumes from last checkpoint.

Download Trained Model

python download_finetuned_model.py \
    --model_name YOUR_USERNAME/fine-tuned-nllb-600M \
    --test

Key Points

• ✅ Checkpoints saved to Hub every 500 steps (~10 min)
• ✅ No ephemeral storage issues
• ✅ No Gradio timeout
• ✅ Resume from checkpoint if interrupted
• ✅ 2-3 hours for standard config (much faster than 8-12 hrs)

Support

Full guide: DOCKER_DEPLOYMENT.md