Pre-download all models: BART, DeBERTa, DistilBART

Models now pre-downloaded in container image:
- Zero-shot models:
* facebook/bart-large-mnli (~1.6GB)
* MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli (~350MB)
* valhalla/distilbart-mnli-12-3 (~500MB)

- Fine-tuning models (tokenizers only):
* microsoft/deberta-v3-small (~50MB)
* microsoft/deberta-v3-base (~100MB)

Total container image: ~4-5GB
/data storage: Only DB + user-trained models (<5GB)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Dockerfile

@@ -29,10 +29,20 @@ RUN mkdir -p /data/.cache/huggingface && chmod -R 777 /data/.cache
 # Create model cache in container (not in /data) to save persistent storage
 RUN mkdir -p /app/.cache && chmod -R 777 /app/.cache
 
-# Pre-download model into container image to avoid using /data storage
+# Pre-download models into container image to avoid using /data storage
 ENV HF_HOME=/app/.cache/huggingface
 ENV TRANSFORMERS_CACHE=/app/.cache/huggingface
-RUN python -c "from transformers import pipeline; pipeline('zero-shot-classification', model='facebook/bart-large-mnli', device=-1)"
+
+# Download zero-shot models (for analysis)
+RUN python -c "from transformers import pipeline; \
+    pipeline('zero-shot-classification', model='facebook/bart-large-mnli', device=-1); \
+    pipeline('zero-shot-classification', model='MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli', device=-1); \
+    pipeline('zero-shot-classification', model='valhalla/distilbart-mnli-12-3', device=-1)"
+
+# Download common fine-tuning models
+RUN python -c "from transformers import AutoModelForSequenceClassification, AutoTokenizer; \
+    AutoTokenizer.from_pretrained('microsoft/deberta-v3-small'); \
+    AutoTokenizer.from_pretrained('microsoft/deberta-v3-base')"
 
 # Hugging Face Spaces uses port 7860
 EXPOSE 7860