comoe-artifacts / README.md
jaygala24's picture
rename ckpts recognizably; remove mislabeled Instruct codec (exp6-iso206x); add README
e5f9cd1 verified
|
Raw
History Blame Contribute Delete
1.18 kB
# comoe-artifacts
Artifacts for **communication-efficient-moe**: dispatch all-to-all compression via
learned linear codecs on a frozen Qwen3-30B-A3B, at 2x.
> **Base model:** everything here pairs with the **SFT'd base**
> (`models/sft-base-qwen3-30b-a3b/`, no-codec 5-task avg **81.7**). A previously
> uploaded `exp6-iso206x` codec was for `Qwen/Qwen3-30B-A3B-Instruct-2507` (a
> *different* model) and has been **removed** to avoid the mismatch.
## Structure
| Path | What |
|---|---|
| `models/sft-base-qwen3-30b-a3b/` | frozen SFT'd base (fullft dolci); the training/eval backbone |
| `offline-codecs/dispatch-2x-iso-init/` | iso2x closed-form warm init |
| `offline-codecs/dispatch-2x-recon010-init/` | iso-recon@0.10 warm init (used by session configs) |
| `offline-codecs/dispatch-2x-func-realloc-init/` | Stage B functional-reallocated warm init |
| `data/raw/`, `data/tokenized/` | dolci-no-tools (raw arrow + tokenized qwen3-2048) |
| `activation_cache/dolci-500k-s1234-step4000/` | dispatch activation cache (for building codecs) |
| `evals/` | lm-eval JSON results per run |
Codecs are the `{compression.pt, meta.json}` format the train/eval scripts expect.