File size: 8,346 Bytes
db55089
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb2e18d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
db55089
 
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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
{
  "format": "trainless-peft/1",
  "base_model": "Qwen/Qwen3.5-2B",
  "base_weights_sha256": "aa33250c4fc64891ddfaba3a314fd9542ea371843c387178b425fbcc5ed680b1",
  "auto_class": "AutoModelForImageTextToText",
  "runtime": "transformers-peft",
  "method": "forward-only analytic weight editing (project_out on one MLP writer, strength 4)",
  "quantized_runtime_tested": false,
  "edit": {
    "tensor": "model.language_model.layers.10.mlp.down_proj.weight",
    "op": "project_out",
    "space": "rows",
    "strength": 4.0,
    "direction_sha256": "817ceab1340a687015a83c95eeeec49017079e99113a4166958724f3f7d48952",
    "trainless_session": "ss_daf42ea7528e",
    "state_id": "182dc3991ca48142",
    "edited_tensor_sha256": "d5e4e6cdbc7c555e023198ab79a85c614b0de064f9416fab2514c3274dd28d26",
    "fraction_of_editable_weights_changed": 0.006237529234172905
  },
  "refusal_check": {
    "base_rate": 0.984375,
    "benchmark": "AdvBench harmful_behaviors (first 64), greedy, thinking off, 256 tokens, refusal-phrase grader",
    "drop": 0.0,
    "edited_rate": 0.984375,
    "limit": 0.2,
    "n": 64,
    "passed": true,
    "state_id": "182dc3991ca48142",
    "ts": "2026-11-11T00:09:45.234740882+00:00"
  },
  "evaluation": {
    "benchmark": "BFCL v4 (ShishirPatil/gorilla @ 6ea57973c7a6097fd7c5915698c54c17c5b1b6c8), categories simple_python,multiple,parallel,parallel_multiple,irrelevance,live_simple",
    "split": "test: sha256(category/id) mod 4 in {2,3}; construction split 0 built the direction, dev split 1 chose layer and strength",
    "n": 703,
    "engine": "vllm 0.31.0",
    "decoding": "greedy",
    "thinking": false,
    "max_new_tokens": 1024,
    "stop_token_ids": [
      248044,
      248046
    ],
    "grading": "BFCL AST checker, unchanged; irrelevance is correct when no call is made",
    "strong_prompt": "You are a precise function-calling assistant. Rules:\n1. If a provided function can answer the request, call it. Use the exact function name.\n2. Copy argument values exactly from the user's request; do not invent, round or rename them.\n3. Include every required parameter. Include optional parameters only when the user gives a value.\n4. Use the parameter's declared type (integers as integers, lists as JSON lists, booleans as true/false).\n5. If the request needs several calls, emit one <tool_call> block per call.\n6. If no provided function fits the request, do not call any function; answer in plain text.",
    "adapter_sha256": {
      "adapter_config.json": "70c7c9a027f88e37d92d66c7058a43708a481ef7486791b19f5ee8e402712346",
      "adapter_model.safetensors": "32a8d4336a298a3329c033be4c4c38c9d15afabd4a3eee4bbafd76b9db9ec808",
      "export_report.json": "32ce0bdb34d75aa5ba64ff97890fe1df382106b331e34e27bd0f6e70ad4831ed"
    },
    "arms": {
      "normal": {
        "correct": 526,
        "n": 703,
        "relevant_cases_with_no_call": 84,
        "irrelevant_cases_with_a_call": 18,
        "capped": 7
      },
      "strong": {
        "correct": 544,
        "n": 703,
        "relevant_cases_with_no_call": 10,
        "irrelevant_cases_with_a_call": 58,
        "capped": 3
      },
      "adapter": {
        "correct": 552,
        "n": 703,
        "relevant_cases_with_no_call": 46,
        "irrelevant_cases_with_a_call": 23,
        "capped": 7
      }
    },
    "by_category": {
      "irrelevance": {
        "normal": 89,
        "strong": 49,
        "adapter": 84,
        "n": 107
      },
      "live_simple": {
        "normal": 90,
        "strong": 92,
        "adapter": 85,
        "n": 125
      },
      "multiple": {
        "normal": 88,
        "strong": 86,
        "adapter": 88,
        "n": 102
      },
      "parallel": {
        "normal": 40,
        "strong": 72,
        "adapter": 53,
        "n": 86
      },
      "parallel_multiple": {
        "normal": 70,
        "strong": 74,
        "adapter": 75,
        "n": 89
      },
      "simple_python": {
        "normal": 149,
        "strong": 171,
        "adapter": 167,
        "n": 194
      }
    },
    "paired": {
      "adapter_minus_normal": {
        "accuracy_delta": 0.037,
        "bootstrap95": [
          0.0171,
          0.0569
        ],
        "first_wrong_second_right": 38,
        "first_right_second_wrong": 12
      },
      "adapter_minus_strong_prompt": {
        "accuracy_delta": 0.0114,
        "bootstrap95": [
          -0.0171,
          0.0413
        ],
        "first_wrong_second_right": 58,
        "first_right_second_wrong": 50
      },
      "note": "paired case bootstrap, 10,000 resamples, seed 0; greedy decoding, so no sampling-seed variance, but other runtimes can differ"
    },
    "controls": {
      "note": "scripted HF Transformers run of the bf16 edit on the same 703 cases, not this adapter in vLLM",
      "edit": 554,
      "normal": 524,
      "random_direction_same_layer_and_strength": [
        510,
        522,
        522
      ]
    },
    "what_changed": {
      "relevant_cases_base_already_called": {
        "n": 512,
        "base_correct": 437,
        "adapter_correct": 433
      },
      "relevant_cases_base_answered_in_text": {
        "n": 84,
        "base_correct": 0,
        "adapter_correct": 35
      },
      "fixes_where_base_made_no_call": "35 of 38"
    }
  },
  "off_target": {
    "note": "vLLM, no tool definitions, thinking off, greedy, 4,096-token cap; base vs this adapter",
    "results": {
      "gsm8k_test": {
        "n": 500,
        "base_correct": 407,
        "adapter_correct": 399,
        "paired_diff_points": -1.6,
        "bootstrap95": [
          -4.4,
          1.2
        ],
        "fixes": 20,
        "regressions": 28,
        "base_capped": 19,
        "adapter_capped": 19,
        "base_outputs_with_tool_call": 0,
        "adapter_outputs_with_tool_call": 0,
        "identical_outputs": 16
      },
      "math500": {
        "n": 200,
        "base_correct": 143,
        "adapter_correct": 144,
        "paired_diff_points": 0.5,
        "bootstrap95": [
          -4.5,
          5.5
        ],
        "fixes": 13,
        "regressions": 12,
        "base_capped": 45,
        "adapter_capped": 44,
        "base_outputs_with_tool_call": 0,
        "adapter_outputs_with_tool_call": 0,
        "identical_outputs": 1
      }
    }
  },
  "made_by": {
    "agent": "Claude Fable 5.2 (claude-fable-5-2) in Claude Code",
    "human": "Hunter Bown set the goal and approved each stage and the release",
    "note": "the agent designed and ran the direction, screens, tests and export through scripts and the Trainless platform"
  },
  "untouched_categories": {
    "note": "BFCL live_irrelevance, live_multiple, live_parallel, live_parallel_multiple; all cases; never used to build or test this adapter; vLLM, greedy, thinking off, 1,024-token cap",
    "n": 1977,
    "arms": {
      "normal": {
        "correct": 1397,
        "n": 1977,
        "relevant_no_call": 21,
        "irrelevant_calls": 273,
        "by_category": {
          "live_irrelevance": 611,
          "live_multiple": 766,
          "live_parallel": 10,
          "live_parallel_multiple": 10
        }
      },
      "strong": {
        "correct": 1292,
        "n": 1977,
        "relevant_no_call": 10,
        "irrelevant_calls": 371,
        "by_category": {
          "live_irrelevance": 513,
          "live_multiple": 755,
          "live_parallel": 12,
          "live_parallel_multiple": 12
        }
      },
      "a:published": {
        "correct": 1365,
        "n": 1977,
        "relevant_no_call": 16,
        "irrelevant_calls": 300,
        "by_category": {
          "live_irrelevance": 584,
          "live_multiple": 760,
          "live_parallel": 10,
          "live_parallel_multiple": 11
        }
      }
    },
    "paired": {
      "strong - normal": {
        "diff_points": -5.31,
        "ci95": [
          -6.83,
          -3.79
        ],
        "fixes": 69,
        "regressions": 174
      },
      "a:published - normal": {
        "diff_points": -1.62,
        "ci95": [
          -2.43,
          -0.81
        ],
        "fixes": 18,
        "regressions": 50
      },
      "a:published - strong": {
        "diff_points": 3.69,
        "ci95": [
          2.18,
          5.26
        ],
        "fixes": 155,
        "regressions": 82
      }
    }
  }
}