Update moe guide

moe/README.md

@@ -14,6 +14,25 @@ multi-token prediction heads.
 - Top-2 routing during training.
 - Shared expert for common syntax and indentation.
 - Router z-loss and auxiliary load-balancing loss.
+- QK norm for long-context attention stability.
+- Expert-choice telemetry from the first step, not after loss diverges.
+- Optional multi-token prediction heads with low loss weight.
+
+## Best Current Shape
+
+Use shared attention and sparse FFN. Keep early layers mostly dense/shared so
+the model learns lexical syntax before specializing.
+
+- Early blocks: dense FFN or shared expert heavy.
+- Middle blocks: sparse MoE every other block.
+- Late blocks: sparse MoE with stricter router stability checks.
+- Expert count: start with 8 experts for smoke, then 16-32 for H100 ablation.
+- Routing: top-2 with capacity factor 1.25 during training.
+- Inference ablation: top-1 only after top-2 training is stable.
+
+For code completion, the highest-value specialization is not one expert per
+language. Better targets are syntax/indentation, FIM reconstruction,
+repository-context patterns, generated-code continuation, and real-code APIs.
 
 ## Routing Metrics
 
@@ -32,3 +51,5 @@ Log:
 - Router churn.
 - Synthetic Python overfitting.
 - MTP loss overpowering next-token loss.
+- Shared expert absorbing too much load.
+- Dropped tokens above 1% for sustained windows.

moe/TRAINING_PLAN.md

@@ -5,6 +5,16 @@
 Train only on manifest-verified JSONL checkpoints. Do not point the trainer at
 raw source directories.
 
+Primary input:
+
+```text
+dataset/<checkpoint>/dataset/*.jsonl
+```
+
+Each 20 GiB checkpoint should be treated as an atomic training unit. Keep the
+checkpoint report beside the run metadata so loss curves can be traced back to
+the exact JSONL files.
+
 ## Curriculum
 
 1. Warm up on short and medium FIM examples.
@@ -12,7 +22,7 @@ raw source directories.
 3. Add code generation as an auxiliary continuation task.
 4. Fine-tune with real-code FIM weighted higher than synthetic continuation.
 
-## Defaults
+## H100 Defaults
 
 - Context: start at 2048, then ablate 4096.
 - Precision: bf16 on H100.
@@ -20,3 +30,26 @@ raw source directories.
 - Grad clipping: enabled.
 - MoE aux loss: start small and tune from expert load.
 - Top-k: top-2 for training, top-1 as inference ablation only.
+- Capacity factor: 1.25 for training.
+- Router z-loss: enabled from step 0.
+- Router jitter: small nonzero value during early training.
+- MTP loss: optional, low coefficient, disabled if next-token validation
+  worsens.
+
+## Preprocessing
+
+- Preserve whitespace and FIM markers exactly.
+- Append `<|endoftext|>` between packed records.
+- Bucket by token length before batching.
+- Track domain and language metadata for router analysis.
+- Do not rebalance by duplicating JSONL files. Rebalance in the dataloader or
+  sampler to avoid SSD growth.
+
+## Router Guardrails
+
+- If one expert holds more than 35% sustained load, raise aux loss or increase
+  shared capacity.
+- If dropped tokens exceed 1%, reduce batch tokens or increase capacity.
+- If router entropy collapses early, lower LR or increase warmup.
+- If domain-to-expert correlation is too strong, improve data mix before adding
+  experts.