# dispatch-2x-sftbase-e2e-step7000

**Best e2e-trained dispatch-2x codec for the SFT'd base** (`models/sft-base-qwen3-30b-a3b`).

- Run: `exp6obj-alltokenkl-lmresp-sft-8k-dp8-zgbfix` (all-token-KL objective,
  obj = lm0.25 + kl + recon + router_kl, lr 1e-5, zero-grad-buffer fixed, 8000 steps).
- Checkpoint: **step-7000**. Load `compression.pt` + `meta.json`; pairs with the SFT base.

Eval (5-task, %):

| | GSM8K | MMLU | HellaSwag | ARC-e | ARC-c | AVG |
|---|---|---|---|---|---|---|
| this codec (step-7000) | 85.4 | 76.5 | 77.5 | 87.0 | 66.6 | **78.6** |
| no-codec (SFT base) | 86.8 | 81.3 | 82.0 | 89.4 | 68.4 | 81.6 |
| gap | -1.4 | -4.8 | -4.5 | -2.4 | -1.8 | **-3.0** |

Stable run (avg 78.0 -> 78.6 from step 1000 on, no collapse). This is the best
verified dispatch-2x on the SFT base. NOTE: the separate `exp6-iso206x` (avg 78.8)
was for `Qwen3-30B-A3B-Instruct-2507`, a different model, and is not in this repo.