Add evaluation metrics for bonus3-lora-moe

README.md

@@ -1,51 +1,13 @@
-# Bonus 3: LoRA for MoE Experts
 
-## Model
-
-Parameter-efficient fine-tuning of Mixture-of-Experts using **LoRA (Low-Rank Adaptation)**.
-
-## Architecture
-
-- 4 transformer layers with MoE
-- 8 experts per layer
-- Top-2 routing
-- LoRA rank: 16, alpha: 32
-
-## Parameter Efficiency
-
-- **Total Parameters**: 55,228,676
-- **Trainable (LoRA)**: 21,625,092 (39.16%)
-- **Frozen (Base)**: 33,603,584 (60.84%)
-- **Reduction**: 2.6x fewer trainable parameters
-
-## Performance
-
-- **Validation Accuracy**: 0.6400
-- **Dataset**: XSum (topic classification)
-- **Training Samples**: 5,000
-
-## LoRA Benefits
-
-1. **Memory Efficient**: Only store small adapter matrices
-2. **Fast Training**: Fewer parameters to update
-3. **Task Switching**: Swap LoRA adapters for different tasks
-4. **Merge Friendly**: Can merge adapters back into base weights
-
-## Files
-
-- `model.pt`: Full model checkpoint
-- `lora_adapters.pt`: Only LoRA parameters (smaller file)
-- `metrics.json`: Training metrics and config
-- `history.csv`: Training history
-
-## Usage
-
-```python
-# Load full model
-checkpoint = torch.load('model.pt')
-model.load_state_dict(checkpoint['model_state_dict'])
-
-# Or load only LoRA adapters (requires base model)
-lora_checkpoint = torch.load('lora_adapters.pt')
-model.load_state_dict(lora_checkpoint['lora_state_dict'], strict=False)
-```
+# Bonus 3: LoRA MoE (XSum)
+
+## Metrics
+- ROUGE-1: 0.0000  
+- ROUGE-2: 0.0000  
+- ROUGE-L: 0.0000  
+- ROUGE-Lsum: 0.0000  
+- SacreBLEU: 0.0000  
+- BERTScore (P/R/F1): 0.0000 / 0.0000 / 0.0000  
+- Compression ratio: 0.0000  
+- Extractiveness: 0.0000  
+- NLI factual consistency: 0.0000