Hugging Face's logo

Snapkitty
/
tensor-roll

snapkitty
october-2026-drop
python
cuda-q
qsharp
quantization
distillation
model-compression
machine-learning
Model card Files
xet
 Community
tensor-roll
File size: 3,527 Bytes
9e551a1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
135
136
137
138
139
140
141
142
143
144
{
  "protocol": "V0 baseline vs V1 focused-distillation",
  "prompts": [
    [
      3,
      1,
      4,
      1
    ],
    [
      7,
      7,
      2,
      9
    ],
    [
      0,
      5,
      5,
      0
    ]
  ],
  "n_tokens": 24,
  "section6": {
    "n_leaves": 148,
    "baseline_decisions": {
      "QUANTIZE": 65,
      "PRESERVE": 83
    },
    "fidelity_aware_decisions": {
      "QUANTIZE": 55,
      "PRESERVE": 93
    },
    "n_decision_diffs": 12,
    "baseline_cost": 11645.0,
    "fidelity_cost": 11456.0
  },
  "V0": {
    "agreement": "7/72",
    "agree_frac": 0.09722222222222222,
    "mean_kl": 1.547805292394603
  },
  "trace_V0": {
    "agreement": "5/24",
    "first_divergence": {
      "position": 0,
      "teacher_token": 30,
      "student_token": 4,
      "kl_div": 2.7323540578677252,
      "js_div": 0.4399995705020577,
      "first_diverging_layer": "layer_0",
      "layer_hidden_rel_err": 1.0380715313927222,
      "layer_attn_div": 0.1767730414867401,
      "all_layer_hidden_rel_err": {
        "layer_0": 1.0380715313927222
      },
      "above_hidden_thresh": {
        "layer_0": 1.0380715313927222
      }
    },
    "mean_kl": 2.5741920989727904,
    "mean_js": 0.43385868451562576
  },
  "ablation_V0": [
    {
      "component": "baseline(student)",
      "agreement": 0.10416666666666667,
      "mean_kl": 2.337756615069645,
      "delta_agree": 0,
      "n_swapped": 0
    },
    {
      "component": "mlp_down",
      "agreement": 0.08333333333333333,
      "mean_kl": 3.0453086743936697,
      "delta_agree": -0.020833333333333343,
      "delta_kl": 0.7075520593240245,
      "n_swapped": 2
    },
    {
      "component": "normalization",
      "agreement": 0.0625,
      "mean_kl": 3.8407641291009362,
      "delta_agree": -0.04166666666666667,
      "delta_kl": 1.503007514031291,
      "n_swapped": 6
    },
    {
      "component": "mlp_up",
      "agreement": 0.041666666666666664,
      "mean_kl": 3.3130315933929992,
      "delta_agree": -0.0625,
      "delta_kl": 0.9752749783233541,
      "n_swapped": 2
    },
    {
      "component": "embeddings",
      "agreement": 0.020833333333333332,
      "mean_kl": 3.8539798822845035,
      "delta_agree": -0.08333333333333334,
      "delta_kl": 1.5162232672148583,
      "n_swapped": 1
    },
    {
      "component": "attention_projections",
      "agreement": 0.020833333333333332,
      "mean_kl": 3.8425005798316954,
      "delta_agree": -0.08333333333333334,
      "delta_kl": 1.5047439647620502,
      "n_swapped": 6
    },
    {
      "component": "attention_output",
      "agreement": 0.020833333333333332,
      "mean_kl": 3.4268810943999513,
      "delta_agree": -0.08333333333333334,
      "delta_kl": 1.089124479330306,
      "n_swapped": 2
    },
    {
      "component": "output_head",
      "agreement": 0.0,
      "mean_kl": 2.535711207020847,
      "delta_agree": -0.10416666666666667,
      "delta_kl": 0.19795459195120202,
      "n_swapped": 2
    }
  ],
  "V1": {
    "agreement": "2/72",
    "agree_frac": 0.027777777777777776,
    "mean_kl": 1.2794614517510858
  },
  "delta_agree": -0.06944444444444445,
  "delta_kl": -0.26834384064351724,
  "V1b": {
    "agreement": "4/72",
    "agree_frac": 0.05555555555555555,
    "mean_kl": 1.1872056255483905
  },
  "V1b_note": "V1 measured student_distilled.npz without finetune; V1b reruns finetune so V0 vs V1b differ only in the extra hidden/logit distillation stages.",
  "delta_agree_V1b": -0.04166666666666667,
  "delta_kl_V1b": -0.3605996668462126
}