Audit fix: Update docs/KAGGLE_RUN.md for ViuMini-Dense-360M

docs/KAGGLE_RUN.md

@@ -58,9 +58,9 @@ subprocess.run([sys.executable, "-m", "pip", "install", "-q", "bitsandbytes>=0.4
 print("[3/5] Verifying 32-layer Dense Architecture...")
 sys.path.insert(0, os.path.join(REPO_DIR, "model", "scripts"))
 import torch
-import viu_moe
+import viu1_moe
 
-mcfg = viu_moe.ViuArgs(
+mcfg = viu1_moe.ViuArgs(
     dim=960,
     n_layers=32,
     n_heads=15,
@@ -71,7 +71,7 @@ mcfg = viu_moe.ViuArgs(
     tie_embeddings=True,
     qk_norm=True
 )
-test_model = viu_moe.Viu1MoE(mcfg)
+test_model = viu1_moe.Viu1MoE(mcfg)
 pcount = sum(p.numel() for p in test_model.parameters())
 print(f"[arch] ViuMini-Dense-360M parameter count: {pcount:,} (~366.6M)")
 del test_model
@@ -134,17 +134,17 @@ else:
 | Metric | Target / Expected Value |
 |---|---|
 | **GPUs** | 2× NVIDIA Tesla T4 (16GB GDDR6 each) |
-| **Micro-batch size** | 4 per GPU (Total 8 across 2 GPUs) |
-| **Gradient Accumulation** | 16 passes |
-| **Effective Batch Size** | **128 sequences** = **131,072 tokens / step** |
+| **Curriculum Batching** | Stage 1 (seq 512, micro 8) $\rightarrow$ Stage 2 (seq 1024, micro 4) $\rightarrow$ Stage 3 (seq 2048, micro 2) |
+| **Gradient Accumulation** | 8 passes per GPU (effective 16 passes) |
+| **Effective Batch Tokens**| Constant **65,536 tokens / step** across all 3 curriculum stages |
 | **Training Speed** | **~8,500 – 10,000 tokens/sec** |
 | **VRAM Usage** | **~3.4 GB / 16 GB** per GPU (12.6 GB free headroom) |
 | **Optimizer** | `PagedAdamW8bit` (learning rate: $2.5 \times 10^{-4}$, warmup: 1,000) |
 | **Gradient Clipping** | 1.0 |
-| **Initial Loss** | $\approx 10.78$ ($\ln(48,000)$ theoretical random uniform) |
+| **Initial Loss** | $\approx 10.78$ ($\ln(48,000)$ theoretical uniform cross-entropy) |
 | **Loss at Step 1,000** | $\approx 6.8 – 7.2$ |
 | **Loss at Step 5,000** | $\approx 4.8 – 5.2$ |
-| **Checkpoints & HF Sync** | Every **1,000 steps** (~3.7 hours / 131M tokens) pushed directly to Hub |
+| **Checkpoints & HF Sync** | Every **1,000 steps** (~2.14 hours / ~65.5M tokens) pushed directly to Hub |
 
 ---