Fix: Add protobuf dependency for DeBERTa models

- Added protobuf>=3.20.0 to requirements.txt
- Removed fine-tuning model pre-download (downloaded on-demand)
- Keep only zero-shot models in container (BART, DeBERTa, DistilBART)
- Added progress messages during model downloads

This fixes the ImportError for DeBERTa tokenizers

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

Co-Authored-By: Claude <[email protected]>

Dockerfile

@@ -33,16 +33,16 @@ RUN mkdir -p /app/.cache && chmod -R 777 /app/.cache
 ENV HF_HOME=/app/.cache/huggingface
 ENV TRANSFORMERS_CACHE=/app/.cache/huggingface
 
-# Download zero-shot models (for analysis)
+# Download zero-shot models (for immediate analysis capability)
+# These are loaded on first analysis, pre-downloading saves time and /data space
 RUN python -c "from transformers import pipeline; \
+    print('Downloading BART-large-MNLI...'); \
     pipeline('zero-shot-classification', model='facebook/bart-large-mnli', device=-1); \
+    print('Downloading DeBERTa-v3-base-MNLI...'); \
     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')"
+    print('Downloading DistilBART-MNLI...'); \
+    pipeline('zero-shot-classification', model='valhalla/distilbart-mnli-12-3', device=-1); \
+    print('All models downloaded successfully')"
 
 # Hugging Face Spaces uses port 7860
 EXPOSE 7860

requirements.txt

@@ -6,6 +6,7 @@ torch==2.5.0
 sentencepiece>=0.2.0
 gunicorn==21.2.0
 hf-transfer>=0.1.5
+protobuf>=3.20.0
 
 # Fine-tuning dependencies
 peft==0.13.2