Add REAP scores (2026-11-12)
Browse files- 28Bob-Riker.json +0 -0
- README.md +69 -0
28Bob-Riker.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
README.md
ADDED
|
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-sa-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- reap
|
| 5 |
+
- moe
|
| 6 |
+
- expert-pruning
|
| 7 |
+
- mixture-of-experts
|
| 8 |
+
- bob
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# REAP scores: 28Bob "Riker"
|
| 12 |
+
|
| 13 |
+
Riker is one of Bob's own brains: Qwen3.6-35B-A3B cut to 208 experts per layer and healed on Bob's lessons. This page holds its experts, scored and ranked by REAP, and the keep lists that were tested, in **one file**:
|
| 14 |
+
`28Bob-Riker.json`. Use it to run the model with fewer experts **without cutting the model file**. Made for
|
| 15 |
+
[Bob](https://github.com/MouseWurX/Bob), the open-source desktop assistant.
|
| 16 |
+
|
| 17 |
+
Bob and his brains' names are inspired by Dennis E. Taylor's novels (not affiliated).
|
| 18 |
+
|
| 19 |
+
## What REAP is
|
| 20 |
+
|
| 21 |
+
REAP (Router-weighted Expert Activation Pruning, Cerebras 2025) scores each expert by **router weight × the size of the
|
| 22 |
+
expert's output**, summed over every word it was picked for. Here the sums run over **100 of Bob's training lessons
|
| 23 |
+
(270,267 tokens)**, with the model served at 12 experts per word from MouseWurX/28Bob-Riker: 28Bob-Q4_K_M.gguf. Experts that are picked rarely but
|
| 24 |
+
matter a lot when they are keep a high score; counting picks alone would miss them.
|
| 25 |
+
|
| 26 |
+
## The file
|
| 27 |
+
|
| 28 |
+
`28Bob-Riker.json`: `model`, `base`, `method`, `lessons`, `tokens`, `layers` (40), `experts` (208 per layer), `experts_per_word` (12),
|
| 29 |
+
`scores.picks` and `scores.mass` (one row per layer, one number per expert), `experts_by_layer` (each layer's expert ids,
|
| 30 |
+
strongest first), `keep_lists` (by percent kept: one list of expert ids per layer; `null` = every expert on), `results` (the
|
| 31 |
+
held-out score at each step) and `picked`.
|
| 32 |
+
|
| 33 |
+
## Using a keep list (no file cut)
|
| 34 |
+
|
| 35 |
+
Brain Engine (Bob's llama.cpp build with the expert-mask patch) switches experts on and off live:
|
| 36 |
+
|
| 37 |
+
```bash
|
| 38 |
+
# keep only the listed experts in each layer (here the 35% list)
|
| 39 |
+
jq '{keep: .keep_lists["35"]}' 28Bob-Riker.json | curl -X POST http://127.0.0.1:8080/expert-mask -H "Content-Type: application/json" -d @-
|
| 40 |
+
# every expert back on
|
| 41 |
+
curl -X POST http://127.0.0.1:8080/expert-mask -H "Content-Type: application/json" -d '{"keep": null}'
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
A mask changes which experts the router may pick; it doesn't free memory by itself. To save memory, load only the kept
|
| 45 |
+
experts (for example, keep the rest in system RAM), or cut a file from the same list.
|
| 46 |
+
|
| 47 |
+
## Held-out results
|
| 48 |
+
|
| 49 |
+
Judged on Bob's 196 held-out lessons (314 assistant turns, never trained on), thinking off. "Right" means the same tools and
|
| 50 |
+
arguments as the full Qwen3.6-35B-A3B teacher, or correctly no tool.
|
| 51 |
+
|
| 52 |
+
| Kept per layer | Share | Right tool | Right arguments | Correctly no tool | Held-out overall | Tokens/s |
|
| 53 |
+
|---|---|---|---|---|---|---|
|
| 54 |
+
| 208 of 208 | 100% | 70/123 | 47/123 | 163/191 | **210/314** (66.9%) | 114.4 |
|
| 55 |
+
| 156 of 208 | 75% | 66/123 | 45/123 | 159/191 | **204/314** (65.0%) | 94.2 |
|
| 56 |
+
| 104 of 208 | 50% | 64/123 | 32/123 | 131/191 | **163/314** (51.9%) | 97.7 |
|
| 57 |
+
| 73 of 208 | 35% | 66/123 | 33/123 | 39/191 | **72/314** (22.9%) | 92.5 |
|
| 58 |
+
|
| 59 |
+
The calibration's own pick (smallest step within 2 points of the full score): **75%**. Read it with the table: 75% is 6 points down.
|
| 60 |
+
|
| 61 |
+
## Credit
|
| 62 |
+
|
| 63 |
+
Base model: Qwen3.6-35B-A3B by Alibaba's Qwen team (Apache 2.0, https://huggingface.co/Qwen/Qwen3.6-35B-A3B). The scores and keep lists here are CC BY-SA 4.0.
|
| 64 |
+
|
| 65 |
+
## Other REAP scores
|
| 66 |
+
|
| 67 |
+
- [35Bob "Bobert"](https://huggingface.co/datasets/MouseWurX/REAP-scores-35Bob-Bobert)
|
| 68 |
+
- [gpt-oss-120b](https://huggingface.co/datasets/MouseWurX/REAP-scores-gpt-oss-120b)
|
| 69 |
+
- Bob's lessons, and the answers from this benchmark run: [MouseWurX/bob-lessons](https://huggingface.co/datasets/MouseWurX/bob-lessons)
|