File size: 2,386 Bytes
f8ee8dc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | ---
license: mit
pretty_name: DeepSeek-V4-Flash-0731 REAM calibration statistics
tags:
- mixture-of-experts
- pruning
- expert-merging
- deepseek-v4
- interpretability
---
# DeepSeek-V4-Flash-0731 — expert calibration statistics (REAM line)
Layerwise routed-expert statistics of
[`deepseek-ai/DeepSeek-V4-Flash-0731`](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731)
(43 MoE layers × 256 experts), collected by running the full model over a
~4.9M-token multi-domain calibration mix (multi-turn dialogs, thinking and
direct modes, rendered with the model's own chat encoder). These are the
statistics behind the REAM144/96 release line — published so that expert
selection, pruning, merging and routing research can start WITHOUT the
expensive H100 collection pass.
## Files
- `layer_XX.npz` (one per MoE layer):
- `saliency [8, 256]` — per-domain REAP-style saliency
`S_i = f_i · E[gate_i · ‖expert_i(x)‖₂ | i ∈ Top-6]`; domain order in
`domains.json`;
- `freq [256]` — Top-6 selection counts over the mix;
- `coact [256, 256]` — joint Top-6 co-activation counts;
- `load [256]` — historical load share (sums to 1);
- `gsum [256]`, `token_count` — auxiliary.
- `imatrix_raw.npz` — squared-input accumulators per expert
(`gate [43, 256, H]`, `down [43, 256, I]`, `calls [43]`) — the raw material
for llama.cpp-style importance matrices of ANY expert subset or merge
(weighted sums of member rows; see `emit_imatrix_merged.py` in the release
pipeline).
- `domains.json` — domain names (canonical order used by the `saliency`
axis) and mix shares.
- `nll_heldout.json` — the source model's NLL on the held-out slice of the
same mix, collected during the pass (a reference point for compressed
variants).
## What these enable
- Reproducing/improving the published selections (each release carries its
`SELECTION.json`);
- Expert pruning/merging experiments at zero collection cost;
- Routing analyses: language/domain specialization by depth, co-activation
cluster structure, load distributions.
## Collection notes
Single pass, batch 6 × 4096 tokens, deterministic packer; the calibration
texts themselves are NOT included and are not recoverable from these
aggregates. Related models: the REAM144/96 line and its merge variants
(see the collection on this profile).
MIT, following the source model.
|