| # Dense Architecture |
|
|
| The Dense model is the control baseline for this dataset. It should use the |
| same tokenizer, data order, validation split, context length, and optimizer |
| family as MoE so architecture changes can be compared cleanly. |
|
|
| ## Recommended Structure |
|
|
| - Decoder-only Transformer. |
| - RMSNorm. |
| - RoPE positional encoding. |
| - Grouped-query attention for lower KV-cache cost. |
| - SwiGLU feed-forward blocks. |
| - 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 |
|
|
| Track: |
|
|
| - Next-token validation loss. |
| - FIM middle exact match. |
| - FIM edit distance. |
| - Syntax parse rate for languages where parsers are available. |
| - Completion latency and tokens/sec. |
|
|
| If Dense does not learn a checkpoint cleanly, inspect data schema, tokenizer, |
| and objective mix before scaling MoE. |
|
|
