Update README.md

README.md

@@ -120,86 +120,11 @@ If you use Gradia in your research, please cite:
 
 ## 📈 Performance Metrics
 
-### Core Training Metrics
 | Metric | Value | Notes |
 |--------|-------|-------|
 | Training Loss | 7.003514766 | Step 10 (best checkpoint) |
 | Perplexity | ~1095.2 | exp(loss) |
-| Loss Reduction Rate | -0.032 per step | Calculated over 10 steps |
-| Convergence Speed | Early (Step 10) | Best model achieved quickly |
-| Training Stability | 100% | No divergence or NaN events |
-
-### Model Architecture Metrics
-| Metric | Value | Notes |
-|--------|-------|-------|
-| Total Parameters | ~937,500 | Estimated from checkpoint size |
-| Embedding Parameters | ~288,768 | Token + positional embeddings |
-| Transformer Parameters | ~629,472 | 4 layers × ~157K params/layer |
-| Layer Norm Parameters | ~19,260 | All normalization layers |
-| Model Size (FP256) | 30MB | Full precision storage |
-| Model Size (FP32 equiv) | 3.75MB | 8x compression potential |
-| Model Size (FP16 equiv) | 1.87MB | 16x compression potential |
-| Parameters per Layer | ~157,368 | Average across 4 transformer layers |
-| Attention Heads per Layer | 8 | 32 dimensions per head |
-
-### FP256 Precision Benefits
-| Metric | Value | Notes |
-|--------|-------|-------|
-| Numerical Stability Saves | 10 | Prevented gradient issues |
-| Extreme Precision Events | 14 | Ultra-precision was crucial |
-| Gradient Stability Improvements | 0 | Raw gradient tracking mode |
-| Training Stability Score | 100, 100, 10... | Consistent high stability |
-| Precision Bits | 256 | vs 32 (FP32) or 16 (FP16) |
-| Decimal Precision | ~77 digits | vs ~7 (FP32) or ~4 (FP16) |
-| Dynamic Range | 2^262143 | Vastly exceeds standard formats |
-| Underflow Prevention Rate | 100% | No gradient underflow detected |
-
-### Memory and Computational Metrics
-| Metric | Value | Notes |
-|--------|-------|-------|
-| Memory Overhead vs FP32 | 8x | 32 bytes vs 4 bytes per param |
-| Memory Overhead vs FP16 | 16x | 32 bytes vs 2 bytes per param |
-| Storage Efficiency | 32 bytes/param | FP256 native storage |
-| Parameter Density | 31,250 params/MB | In FP256 format |
-| Training Memory Peak | ~240MB | Model + gradients + optimizer |
-| Gradient Precision | 256-bit | Full precision gradients |
-
-### Training Efficiency Metrics
-| Metric | Value | Notes |
-|--------|-------|-------|
-| Steps to Best Model | 10 | Early convergence |
-| Training Time per Step | ~1598ms | From checkpoint data |
-| FP256 Update Norm | 0.000316227 | Gradient update magnitude |
-| Learning Rate Precision | 256-bit | Ultra-precise LR updates |
-| Batch Processing Stability | 100% | No batch failures |
-| Optimizer Convergence | Stable | No oscillations detected |
-
-### Comparative Analysis (Estimated)
-| Metric | FP256 (This Model) | FP32 Equivalent | FP16 Equivalent |
-|--------|-------------------|-----------------|-----------------|
-| Model Size | 30MB | 3.75MB | 1.87MB |
-| Precision Digits | ~77 | ~7 | ~4 |
-| Gradient Stability | 10 saves | Likely 2-3 failures | Likely 5-8 failures |
-| Memory Usage | 240MB | 30MB | 15MB |
-| Numerical Range | 2^262143 | 2^127 | 2^15 |
-| Training Stability | 100% | ~85-90% | ~70-80% |
-
-### Vocabulary and Sequence Metrics
-| Metric | Value | Notes |
-|--------|-------|-------|
-| Vocabulary Size | 1,000 | Compact vocab for demo |
-| Max Sequence Length | 128 tokens | Short context window |
-| Embedding Dimension | 256 | Hidden size alignment |
-| Position Embeddings | 128 | Learned positional encoding |
-| Token Coverage | Demo dataset | Limited scope |
-| Sequence Processing | Fixed length | No dynamic batching |
-
-### Experimental Research Metrics
-| Metric | Value | Research Value |
-|--------|-------|----------------|
-| Novel Precision Format | FP256 | First known implementation |
-| Stability Interventions | 10 | Demonstrates precision benefits |
-| Precision Event Rate | 1.4 per step | High precision requirement |
-| Research Reproducibility | Full | Complete checkpoint available |
-| Implementation Novelty | Custom | FP256 transformer architecture |
-| Scientific Contribution | High | Ultra-precision ML exploration |
+| Model Size | 30MB | FP256 precision |
+| Parameters | ~937K | Estimated from checkpoint size |
+| Stability Events | 10 | Numerical instabilities prevented |
+| Precision Events | 14 | Cases where FP256 was crucial |