Update dense guide

dense/README.md

@@ -11,8 +11,23 @@ family as MoE so architecture changes can be compared cleanly.
 - RoPE positional encoding.
 - Grouped-query attention for lower KV-cache cost.
 - SwiGLU feed-forward blocks.
-- Tied embedding/head only after validation.
+- Untied embedding/head for the first quality run; tie only if validation shows
+  no FIM regression.
 - Native support for FIM tokens.
+- QK norm for attention stability at longer context.
+- Residual branch init scaling so deeper runs do not spike early loss.
+
+## Baseline Shape
+
+Use Dense as the clean control model, not as the final scaling target.
+
+- Small smoke: 12-16 layers, 768-1024 hidden, 8-16 heads, context 2048.
+- H100 baseline: 24-32 layers, 2048-3072 hidden, GQA, context 4096.
+- Keep the tokenizer, file order, validation split, optimizer, and batch-token
+  target identical to MoE.
+
+The first Dense run should answer one question: does the dataset/tokenizer learn
+cleanly without router complexity? If not, do not scale MoE yet.
 
 ## Evaluation
 

dense/TRAINING_PLAN.md

@@ -5,6 +5,15 @@
 Use the same JSONL checkpoint bundle as MoE. Keep random seed, batch token
 target, learning-rate schedule, and validation split aligned.
 
+Primary input:
+
+```text
+dataset/<checkpoint>/dataset/*.jsonl
+```
+
+Do not train from hidden staging folders unless rebuilding checkpoints. Hidden
+`data/curated_upload` is a staging queue, not the training contract.
+
 ## Curriculum
 
 1. Start with short and medium records.
@@ -12,8 +21,28 @@ target, learning-rate schedule, and validation split aligned.
 3. Add continuation/code generation records.
 4. Evaluate FIM and next-token loss separately.
 
+## H100 Defaults
+
+- Precision: bf16.
+- Context: 2048 smoke, 4096 main run.
+- Optimizer: AdamW or fused AdamW.
+- LR schedule: warmup then cosine decay.
+- Gradient clipping: enabled.
+- Batch target: choose by total tokens, not examples, because FIM lengths vary.
+- Validation: fixed files or fixed checksums from the checkpoint report.
+
+## Preprocessing
+
+- Preserve whitespace exactly.
+- Append `<|endoftext|>` between records during packing.
+- Keep FIM markers visible; do not strip special tokens.
+- Reject empty `text`, invalid JSONL, and records exceeding context unless
+  deterministic chunking is enabled.
+- Bucket by token length to reduce padding waste.
+
 ## Stop Conditions
 
 - Validation loss diverges twice after LR reduction.
 - FIM metrics regress while train loss improves.
 - Tokenizer round-trip or special-token audit fails.
+- Syntax parse rate drops while raw loss improves.