MouseWurX commited on
Commit
48d4511
·
verified ·
1 Parent(s): a7a317e

28Bob GGUF files, cards, keep list and adapter from Bob's Forge

Browse files
.gitattributes CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ 28Bob-Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
37
+ 28Bob-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
28Bob-Q3_K_M.bob.json CHANGED
@@ -1,58 +1,57 @@
1
- {
2
- "card": 1,
3
- "file": "28Bob-Q3_K_M.gguf",
4
- "bytes": 13873840224,
5
- "name": "28Bob Riker (Q3_K_M)",
6
- "nickname": "Riker",
7
- "family": "Bob",
8
- "madeByBob": true,
9
- "jobs": [
10
- "bob",
11
- "deep",
12
- "tinker",
13
- "reviewer",
14
- "office",
15
- "ledger",
16
- "scout"
17
- ],
18
- "params": 28,
19
- "activeParams": 3,
20
- "vramGB": 15,
21
- "ctx": 262144,
22
- "cutFrom": {
23
- "name": "Qwen3.6 35B-A3B",
24
- "by": "Alibaba",
25
- "hf": "Qwen/Qwen3.6-35B-A3B"
26
- },
27
- "expertsKept": "208 of 256",
28
- "published": "2026-11-08",
29
- "license": "Apache 2.0",
30
- "what": "Bob\u0027s best brain: the full Qwen\u0027s tool habits, coding and exact formats kept, 30% faster; a little vaguer on trivia.",
31
- "benchmark": {
32
- "date": "2026-11-08",
33
- "report": "https://github.com/MouseWurX/Bob/blob/main/docs/FORGE-JOURNAL.md",
34
- "scores": {
35
- "Coding": 69,
36
- "Quick code": 100,
37
- "Tools": 65,
38
- "Facts": 90,
39
- "Routing": 96,
40
- "ARC-Challenge": 92.9,
41
- "ARC-Easy": 97.6,
42
- "ARC-Challenge (thinking)": 93.9,
43
- "ARC-Easy (thinking)": 98.3
44
- },
45
- "baseline": {
46
- "name": "Qwen3.6 35B-A3B",
47
- "scores": {
48
- "Coding": 92,
49
- "Quick code": 100,
50
- "Tools": 67,
51
- "Facts": 100,
52
- "ARC-Challenge": 95.8,
53
- "ARC-Easy": 98.9
54
- }
55
- },
56
- "arcNote": "ARC zero-shot, full test sets (1,172 / 2,376); thinking on with a 1,536-token budget. Baseline thinking-off on full sets."
57
- }
58
  }
 
1
+ {
2
+ "card": 1,
3
+ "file": "28Bob-Q3_K_M.gguf",
4
+ "bytes": 13873840224,
5
+ "sha256": "d3b65bbc3ae8529f721675350b2f2a6e12e3205269345eaf6066d3e535684267",
6
+ "name": "28Bob Riker (Q3_K_M)",
7
+ "nickname": "Riker",
8
+ "family": "Bob",
9
+ "madeByBob": true,
10
+ "jobs": [
11
+ "bob",
12
+ "deep",
13
+ "tinker",
14
+ "reviewer",
15
+ "office",
16
+ "ledger",
17
+ "scout"
18
+ ],
19
+ "params": 29,
20
+ "activeParams": 3,
21
+ "vramGB": 15,
22
+ "cutFrom": {
23
+ "name": "Qwen3.6 35B A3B",
24
+ "by": "Alibaba",
25
+ "hf": "Qwen/Qwen3.6-35B-A3B"
26
+ },
27
+ "published": "2026-11-09",
28
+ "license": "Apache 2.0",
29
+ "what": "A Bob cut from Qwen3.6-35B-A3B for Bob's own jobs: tools, coding, exact formats and everyday chat.",
30
+ "benchmark": {
31
+ "date": "2026-11-08",
32
+ "report": "https://github.com/MouseWurX/Bob/blob/main/docs/FORGE-JOURNAL.md",
33
+ "scores": {
34
+ "Coding": 69,
35
+ "Quick code": 100,
36
+ "Tools": 65,
37
+ "Facts": 90,
38
+ "Routing": 96,
39
+ "ARC-Challenge": 92.9,
40
+ "ARC-Easy": 97.6,
41
+ "ARC-Challenge (thinking)": 93.9,
42
+ "ARC-Easy (thinking)": 98.3
43
+ },
44
+ "baseline": {
45
+ "name": "Qwen3.6 35B-A3B",
46
+ "scores": {
47
+ "Coding": 92,
48
+ "Quick code": 100,
49
+ "Tools": 67,
50
+ "Facts": 100,
51
+ "ARC-Challenge": 95.8,
52
+ "ARC-Easy": 98.9
53
+ }
54
+ },
55
+ "arcNote": "ARC zero-shot, full test sets (1,172 / 2,376); thinking on with a 1,536-token budget. Baseline thinking-off on full sets."
56
+ }
 
57
  }
28Bob-Q3_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d3b65bbc3ae8529f721675350b2f2a6e12e3205269345eaf6066d3e535684267
3
+ size 13873840224
28Bob-Q4_K_M.bob.json CHANGED
@@ -1,58 +1,57 @@
1
- {
2
- "card": 1,
3
- "file": "28Bob-Q4_K_M.gguf",
4
- "bytes": 17494120544,
5
- "name": "28Bob Riker (Q4_K_M)",
6
- "nickname": "Riker",
7
- "family": "Bob",
8
- "madeByBob": true,
9
- "jobs": [
10
- "bob",
11
- "deep",
12
- "tinker",
13
- "reviewer",
14
- "office",
15
- "ledger",
16
- "scout"
17
- ],
18
- "params": 28,
19
- "activeParams": 3,
20
- "vramGB": 19,
21
- "ctx": 262144,
22
- "cutFrom": {
23
- "name": "Qwen3.6 35B-A3B",
24
- "by": "Alibaba",
25
- "hf": "Qwen/Qwen3.6-35B-A3B"
26
- },
27
- "expertsKept": "208 of 256",
28
- "published": "2026-11-08",
29
- "license": "Apache 2.0",
30
- "what": "Bob\u0027s best brain: the full Qwen\u0027s tool habits, coding and exact formats kept, 30% faster; a little vaguer on trivia.",
31
- "benchmark": {
32
- "date": "2026-11-08",
33
- "report": "https://github.com/MouseWurX/Bob/blob/main/docs/FORGE-JOURNAL.md",
34
- "scores": {
35
- "Coding": 69,
36
- "Quick code": 100,
37
- "Tools": 65,
38
- "Facts": 90,
39
- "Routing": 96,
40
- "ARC-Challenge": 92.9,
41
- "ARC-Easy": 97.6,
42
- "ARC-Challenge (thinking)": 93.9,
43
- "ARC-Easy (thinking)": 98.3
44
- },
45
- "baseline": {
46
- "name": "Qwen3.6 35B-A3B",
47
- "scores": {
48
- "Coding": 92,
49
- "Quick code": 100,
50
- "Tools": 67,
51
- "Facts": 100,
52
- "ARC-Challenge": 95.8,
53
- "ARC-Easy": 98.9
54
- }
55
- },
56
- "arcNote": "ARC zero-shot, full test sets (1,172 / 2,376); thinking on with a 1,536-token budget. Baseline thinking-off on full sets."
57
- }
58
  }
 
1
+ {
2
+ "card": 1,
3
+ "file": "28Bob-Q4_K_M.gguf",
4
+ "bytes": 17494120544,
5
+ "sha256": "c96ae59c62b6e1802788d07e2d3e2edb540b3a765dd71689d92a1b546e4439e7",
6
+ "name": "28Bob Riker (Q4_K_M)",
7
+ "nickname": "Riker",
8
+ "family": "Bob",
9
+ "madeByBob": true,
10
+ "jobs": [
11
+ "bob",
12
+ "deep",
13
+ "tinker",
14
+ "reviewer",
15
+ "office",
16
+ "ledger",
17
+ "scout"
18
+ ],
19
+ "params": 29,
20
+ "activeParams": 3,
21
+ "vramGB": 19,
22
+ "cutFrom": {
23
+ "name": "Qwen3.6 35B A3B",
24
+ "by": "Alibaba",
25
+ "hf": "Qwen/Qwen3.6-35B-A3B"
26
+ },
27
+ "published": "2026-11-09",
28
+ "license": "Apache 2.0",
29
+ "what": "A Bob cut from Qwen3.6-35B-A3B for Bob's own jobs: tools, coding, exact formats and everyday chat.",
30
+ "benchmark": {
31
+ "date": "2026-11-08",
32
+ "report": "https://github.com/MouseWurX/Bob/blob/main/docs/FORGE-JOURNAL.md",
33
+ "scores": {
34
+ "Coding": 69,
35
+ "Quick code": 100,
36
+ "Tools": 65,
37
+ "Facts": 90,
38
+ "Routing": 96,
39
+ "ARC-Challenge": 92.9,
40
+ "ARC-Easy": 97.6,
41
+ "ARC-Challenge (thinking)": 93.9,
42
+ "ARC-Easy (thinking)": 98.3
43
+ },
44
+ "baseline": {
45
+ "name": "Qwen3.6 35B-A3B",
46
+ "scores": {
47
+ "Coding": 92,
48
+ "Quick code": 100,
49
+ "Tools": 67,
50
+ "Facts": 100,
51
+ "ARC-Challenge": 95.8,
52
+ "ARC-Easy": 98.9
53
+ }
54
+ },
55
+ "arcNote": "ARC zero-shot, full test sets (1,172 / 2,376); thinking on with a 1,536-token budget. Baseline thinking-off on full sets."
56
+ }
 
57
  }
28Bob-Q4_K_M.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c96ae59c62b6e1802788d07e2d3e2edb540b3a765dd71689d92a1b546e4439e7
3
+ size 17494120544
README.md CHANGED
@@ -1,104 +1,44 @@
1
  ---
2
  license: apache-2.0
3
  base_model: Qwen/Qwen3.6-35B-A3B
4
- tags:
5
- - bob
6
- - gguf
7
- - mixture-of-experts
8
- - expert-pruning
9
- - distillation
10
- - tool-use
11
- language:
12
- - en
13
  ---
14
 
15
  # 28Bob Riker
16
 
17
- Riker is a 28-billion-parameter mixture-of-experts model (about 3 billion active per word) made for the desktop
18
- assistant [Bob](https://github.com/MouseWurX/Bob). He was cut from
19
- [Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) by keeping, in every layer, the 208 of 256 experts that
20
- Bob's own jobs use. Then he was healed with a LoRA for one full pass against the full model's next-word odds, on the
21
- owner's lesson set ([bob-lessons](https://huggingface.co/datasets/MouseWurX/bob-lessons), about 9,000 of Bob's
22
- conversations) plus general text. He was finished on 7 November 2026. Method, numbers and logs: the
23
- [Forge journal](https://github.com/MouseWurX/Bob/blob/main/docs/FORGE-JOURNAL.md).
24
 
25
- He is about 20% smaller than the model he came from, 34% faster on the owner's 16 GB card, and keeps its long-context
26
- reading. He loses a little on facts and harder coding.
 
27
 
28
  ## Files
29
 
30
- | File | Size | For |
31
  |---|---|---|
32
- | `28Bob-Q4_K_M.gguf` | 17.5 GB | 24 GB cards, or 16 GB cards with some layers on system memory |
33
- | `28Bob-Q3_K_M.gguf` | 13.9 GB | 16 GB cards |
34
- | `28Bob-adapter.pt` | 1.7 GB | the healing LoRA; with the keep list it rebuilds the full-precision model from Qwen3.6-35B-A3B |
35
- | `28Bob-keep.json` | 38 KB | the 208 experts kept in each of the 40 layers |
36
- | `coverage-208.md` | small | how the keep list was chosen, and each skill's kept share |
37
- | `heal-report.json` | small | the heal settings and before/after losses |
38
- | `28Bob-Q4_K_M.bob.json`, `28Bob-Q3_K_M.bob.json` | small | Bob's brain cards |
39
 
40
- The GGUF files are being rebuilt from the adapter on a rented server and arrive here next; until then, rebuild from Qwen3.6-35B-A3B with the keep list and adapter.
41
 
42
- Vision: use the Qwen3.6-35B-A3B `mmproj` file; the hidden size is unchanged.
43
 
44
- ## How to run
45
-
46
- With [llama.cpp](https://github.com/ggml-org/llama.cpp) (any recent build with Qwen3.6 MoE support):
47
-
48
- ```
49
- llama-server -m 28Bob-Q4_K_M.gguf -c 32768 --jinja -fa on -ctk q8_0 -ctv q8_0 --fit on
50
- ```
51
-
52
- - 8 experts per word (the file's default; don't lower it).
53
- - 32K context or more. The needle test below reads cleanly at 128K.
54
- - Thinking on or off per request, with the Qwen3.6 chat template (`enable_thinking`).
55
- - On a 16 GB card, `--fit on` (or `--n-cpu-moe`) puts part of the experts on system memory.
56
-
57
- ## Heal (6 to 7 November 2026, one H200)
58
-
59
- | | Before | After | Full Qwen |
60
- |---|---|---|---|
61
- | Held-out lessons, cross-entropy | 0.401 | 0.362 | 0.324 |
62
- | Held-out lessons, KD (distance from the full model) | 0.087 | 0.067 | |
63
- | General text, cross-entropy | 2.154 | 2.158 | 2.101 |
64
- | General text, KD | 0.178 | 0.132 | |
65
-
66
- 929 of 929 steps, 5.9 hours, 63.4 million tokens. LoRA rank 32 (8 on the experts), learning rate 1e-4, one epoch.
67
-
68
- ## Benchmarks (8 November 2026, Q4_K_M, 8 experts per word, the owner's 16 GB card)
69
-
70
- | Test | Riker | Full Qwen3.6-35B-A3B (Q4) |
71
  |---|---|---|
72
- | ARC-Challenge, thinking off (1,172 questions) | **92.9%** | 95.8% |
73
- | ARC-Easy, thinking off (2,376 questions) | **97.6%** | 98.9% |
74
- | ARC-Challenge, thinking on (1,172 questions) | **93.9%** | 97.7% (300-question sample; Riker 96.3% on the same 300) |
75
- | ARC-Easy, thinking on (2,376 questions) | **98.3%** | 99.0% (300-question sample; Riker 97.3% on the same 300) |
76
- | Coding race, thinking on / off (of 13 tasks) | 9 / 7 | 12 / 11 |
77
- | Quick code tasks (of 6) | 6 | 6 |
78
- | Tool score, held-out conversations | 65% | 67% |
79
- | Routing race, Bob's jobs (of 25) | 24 | |
80
- | Facts (of 20) | 18 | 20 |
81
- | Needle in a long document, 8K to 128K (of 12) | 12 (Q3_K_M) | 12 |
82
- | Speed, tokens a second | 111 | 83 |
83
-
84
- ARC is zero-shot: the question and lettered choices, then "Answer with the letter only.", scored on the first letter.
85
- Thinking off uses temperature 0. Thinking on gets a 1,536-token thinking budget and is then asked to answer.
86
-
87
- The other tests are the project's own, run on one machine, for comparing Bobs with each other and with the model they
88
- came from. They are not leaderboard numbers.
89
-
90
- ## Honest notes
91
-
92
- - Facts: 18 of 20 where the full model gets 20. Trivia recall is the main thing the cut cost.
93
- - The keep list was built with the plain 95% coverage rule for every skill. The stronger 99% rule for tool calls,
94
- coding and exact formats, planned for later Bobs, was not used for Riker. Harder coding is where he trails most
95
- (9 of 13 against 12).
96
- - The heal used the owner's own lesson set, so Riker leans towards Bob's jobs: tool calls, routing, coding help and
97
- office tasks.
98
 
99
  ## Licence and credit
100
 
101
- Apache 2.0, the same as the base model. The base model, Qwen3.6-35B-A3B, is by Alibaba's Qwen team. The lesson set he
102
- was healed on is CC BY-SA 4.0, plus an Apache 2.0 subset. The name "Riker" and
103
- the other Bob names are a tribute to Dennis E. Taylor's novels. This project is not affiliated with the author or with
104
- Alibaba.
 
1
  ---
2
  license: apache-2.0
3
  base_model: Qwen/Qwen3.6-35B-A3B
4
+ tags: [bob, gguf, mixture-of-experts, tool-use]
 
 
 
 
 
 
 
 
5
  ---
6
 
7
  # 28Bob Riker
8
 
9
+ A Bob cut from Qwen3.6-35B-A3B for Bob's own jobs: tools, coding, exact formats and everyday chat.
 
 
 
 
 
 
10
 
11
+ Made by [Bob's Forge](https://github.com/MouseWurX/Bob/blob/main/docs/FORGE-JOURNAL.md) from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
12
+ (all experts kept per layer, about 29 B parameters, 3 B active per word).
13
+ Method, numbers and logs are in the Forge journal.
14
 
15
  ## Files
16
 
17
+ | File | Size | Needs |
18
  |---|---|---|
19
+ | `28Bob-Q3_K_M.gguf` | 13.9 GB | about 15 GB of card memory |
20
+ | `28Bob-Q4_K_M.gguf` | 17.5 GB | about 19 GB of card memory |
 
 
 
 
 
21
 
22
+ Each GGUF has a `.bob.json` card beside it that Bob's catalogue reads when you add the file as your own brain.
23
 
24
+ ## Home benchmarks (2026-11-08)
25
 
26
+ | Test | 28Bob | Qwen3.6 35B-A3B |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
27
  |---|---|---|
28
+ | Coding | 69 | 92 |
29
+ | Quick code | 100 | 100 |
30
+ | Tools | 65 | 67 |
31
+ | Facts | 90 | 100 |
32
+ | Routing | 96 | |
33
+ | ARC-Challenge | 92.9 | 95.8 |
34
+ | ARC-Easy | 97.6 | 98.9 |
35
+ | ARC-Challenge (thinking) | 93.9 | |
36
+ | ARC-Easy (thinking) | 98.3 | |
37
+
38
+ The project's own tests on one machine, for comparing Bobs with each other and with the model they came from, not for leaderboards.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
 
40
  ## Licence and credit
41
 
42
+ Apache 2.0, as the base model (Alibaba's Qwen team). The lessons it learned are Bob's own lesson sets (CC BY-SA 4.0) plus
43
+ openly licensed traces from other models, credited in the Forge journal. The Bob names are a tribute to Dennis E. Taylor's novels;
44
+ the project is not affiliated with the author.