| # dispatch-fullrank-sftbase-e2e-step1000 |
|
|
| Full-rank (1x, **no compression**) dispatch codec, e2e-trained on the SFT base |
| model. This is the **lossless sanity codec**: it proves the dispatch codec |
| recipe / objective / pipeline is clean, so any downstream gap at 2x is genuinely |
| a rank/capacity effect, not a pipeline bug. |
|
|
| ## What it is |
|
|
| - **compression** `dispatch`, `compression_ratio: 1` (per-layer bottleneck = full |
| hidden dim 2048), `init: identity_truncated`, `stale_stride: 12`. |
| - **base model** `models/sft-base-qwen3-30b-a3b` (fullft-qwen3base-bf16 step-4000 |
| backbone-hf). Do **not** load against Qwen3-30B-A3B-Instruct-2507. |
| - **training** e2e, frozen backbone, objective `lm0.25 + recon(KL) + router-KL`, |
| bs 64, lr 1e-5, seed 42, DP4 (world_size 16), **step-1000**. |
| |
| ## Eval (lm-eval-harness, same protocol as the 2x codecs) |
| |
| | task | metric | full-rank (this) | no-codec target | |
| |---|---|---|---| |
| | gsm8k | exact_match strict | 86.8 | 86.9 | |
| | mmlu | acc | 81.3 | 80.7 | |
| | hellaswag | acc_norm | 82.1 | 82.5 | |
| | arc_easy | acc_norm | 89.5 | 89.3 | |
| | arc_challenge | acc_norm | 68.9 | 69.1 | |
| | **AVG** | | **81.7** | **81.6** | |
| |
| Lossless within noise (gap ~0). Contrast the best 2x dispatch codec |
| (`trained-codecs/dispatch-2x-sftbase-e2e-step7000`, avg 78.6, gap -3.0). |
| |
| ## Files |
| |
| - `compression.pt` — codec state dict (load as init directly, no transform). |
| - `meta.json` — training provenance (git commit, base model, compression config). |
| |
