Hugging Face's logo

patrickbdevaney
/
deepseek-v4-flash-0731-REAP-calibrated-draft-head

Safetensors
deepseek_v4
speculative-decoding
draft-model
mtp
jetson
fp8
Model card Files
xet
 Community
deepseek-v4-flash-0731-REAP-calibrated-draft-head
File size: 3,592 Bytes
2c11bfb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
{
  "name": "s3",
  "promoted_utc": "2026-08-12T00:15:33Z",
  "engine_git_rev": "85dbea6cfcd0e15651177135dc60b7b273c41445",
  "engine_tree_dirty": true,
  "base_model": "0xSero/DeepSeek-V4-Flash-0731-REAP",
  "base_source_model": "deepseek-ai/DeepSeek-V4-Flash-0731",
  "base_source_revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb",
  "reap": {
    "original_experts_per_layer": 256,
    "kept_experts_per_layer": 160
  },
  "measurement": {
    "protocol": "8-prompt suite (scripts/prompt_suite.json, one per category), NGEN0>=200 so tau is past the drafter's 128-token sliding window (F92); block 6, adaptK 1.50; clean run; canonical 6-token prompt is a CONTROL, excluded from the mean (F96).",
    "suite_mean_tau": 3.8438,
    "suite_mean_tok_s": 25.5312,
    "base_ar_tok_s": 13.8,
    "n_suite_prompts": 8,
    "per_prompt": [
      {
        "block": 6,
        "prompt": 0,
        "tau": 2.87,
        "ms_tok": 46.9,
        "tok_s": 21.33
      },
      {
        "block": 6,
        "prompt": 1,
        "tau": 3.64,
        "ms_tok": 40.7,
        "tok_s": 24.55
      },
      {
        "block": 6,
        "prompt": 2,
        "tau": 4.2,
        "ms_tok": 36.8,
        "tok_s": 27.16
      },
      {
        "block": 6,
        "prompt": 3,
        "tau": 2.49,
        "ms_tok": 53.1,
        "tok_s": 18.83
      },
      {
        "block": 6,
        "prompt": 4,
        "tau": 3.06,
        "ms_tok": 43.9,
        "tok_s": 22.8
      },
      {
        "block": 6,
        "prompt": 5,
        "tau": 3.94,
        "ms_tok": 41.0,
        "tok_s": 24.39
      },
      {
        "block": 6,
        "prompt": 6,
        "tau": 6.09,
        "ms_tok": 28.5,
        "tok_s": 35.08
      },
      {
        "block": 6,
        "prompt": 7,
        "tau": 3.85,
        "ms_tok": 37.5,
        "tok_s": 26.7
      },
      {
        "block": 6,
        "prompt": 8,
        "tau": 3.48,
        "ms_tok": 40.4,
        "tok_s": 24.74
      }
    ],
    "lossless_gate": "PASS"
  },
  "incumbent_at_promotion": {
    "name": "s1",
    "suite_tau": 3.5762,
    "suite_tok_s": 24.515
  },
  "notes": "",
  "files": [
    {
      "file": "config.json",
      "bytes": 3602,
      "sha256": "de9ea59b8bba4cc49f67b4d6a30cebb148ab465aebc57225b355158477a56c8e"
    },
    {
      "file": "model-00046-of-00048.safetensors",
      "bytes": 2326148920,
      "sha256": "83b75bc919c160062ae09c7a4d24ed1646d4b1ee074d07484f5fc4b3d20047c8"
    },
    {
      "file": "model-00047-of-00048.safetensors",
      "bytes": 2275805696,
      "sha256": "8d44875f651e1b4f29748d8da7e853592dc139e6e122a34e5ca68e8f6021d7fd"
    },
    {
      "file": "model-00048-of-00048.safetensors",
      "bytes": 2408468996,
      "sha256": "938d6f17100a1ea766528f66c7d952004342a8b00ca2be956f75299e25ca7df2"
    },
    {
      "file": "model.safetensors.index.json",
      "bytes": 213205,
      "sha256": "3c6ca0116218bd358ad3b8ca8ab067194585edb187d594aef9f42986ec5fe50f"
    },
    {
      "file": "mtp_trained.safetensors",
      "bytes": 1034643326,
      "sha256": "5dd3ae629d3d504f80ab9de23e47db54a475bd162f93188277b3832efd507038"
    }
  ],
  "promoted": true,
  "trained_tensors_present": true,
  "rebuild": "Loadable head regenerated from mtp_trained.safetensors with tools/build_trained_head.py --base <checkpoint> --trained mtp_trained.safetensors --out <dir>; deterministic, self-checked.",
  "training": {
    "sessions": 3,
    "corpus": "1536 prompts balanced 8 ways (1472 train / 64 hold-out)",
    "chunks": 3,
    "steps": 1472,
    "loss": "0.9736 -> 0.0405 (tail mean 0.4446)"
  }
}