Channel-aware attention-head pruning / results.json
Mechanism failed
1{
2 "seed": 2304,
3 "math": {
4 "content_sink_invariance": [
5 {
6 "m": 2,
7 "predicted_range": 0.0,
8 "observed_range": 0.0
9 },
10 {
11 "m": 3,
12 "predicted_range": 0.0,
13 "observed_range": 0.0
14 },
15 {
16 "m": 7,
17 "predicted_range": 0.0,
18 "observed_range": 0.0
19 },
20 {
21 "m": 15,
22 "predicted_range": 0.0,
23 "observed_range": 1.7763568394002505e-15
24 }
25 ],
26 "equal_content_sink_scaling": [
27 {
28 "m": 2,
29 "predicted": 1.3333333333333333,
30 "observed_mean": 1.3333333333333337,
31 "observed_sd": 3.746373528904865e-15
32 },
33 {
34 "m": 3,
35 "predicted": 1.5,
36 "observed_mean": 1.5,
37 "observed_sd": 2.249784163675333e-15
38 },
39 {
40 "m": 7,
41 "predicted": 1.75,
42 "observed_mean": 1.7500000000000007,
43 "observed_sd": 2.7413292758006402e-15
44 },
45 {
46 "m": 15,
47 "predicted": 1.875,
48 "observed_mean": 1.8750000000000002,
49 "observed_sd": 2.6330782899035054e-15
50 },
51 {
52 "m": 31,
53 "predicted": 1.9375,
54 "observed_mean": 1.9374999999999931,
55 "observed_sd": 8.003480046730369e-14
56 }
57 ],
58 "decomposition_check": [
59 {
60 "m": 2,
61 "corrected_max_relative_error": 2.4254150508628115e-15,
62 "stated_median_relative_error": 0.5909399083273867,
63 "stated_mean_relative_error": 0.5769427836274625
64 },
65 {
66 "m": 3,
67 "corrected_max_relative_error": 1.061658402041578e-15,
68 "stated_median_relative_error": 0.7079299585244985,
69 "stated_mean_relative_error": 0.6756008300678972
70 },
71 {
72 "m": 7,
73 "corrected_max_relative_error": 7.334361550024249e-16,
74 "stated_median_relative_error": 0.8507274949842227,
75 "stated_mean_relative_error": 0.8285722161750025
76 },
77 {
78 "m": 15,
79 "corrected_max_relative_error": 7.64761719038795e-16,
80 "stated_median_relative_error": 0.9272963573350566,
81 "stated_mean_relative_error": 0.9085484025310643
82 },
83 {
84 "m": 31,
85 "corrected_max_relative_error": 5.83734922781258e-16,
86 "stated_median_relative_error": 0.9621593741714742,
87 "stated_mean_relative_error": 0.9538432291282283
88 }
89 ]
90 },
91 "pruning_toy": {
92 "pairs": [
93 {
94 "pair": [
95 0,
96 1
97 ],
98 "names": [
99 "reference",
100 "same_sink_diff_content"
101 ],
102 "cosine": 0.0008771847295698798,
103 "js": 0.018550749134822735,
104 "content": 5.20819598644114,
105 "sink": 0.0,
106 "full": 26.040979932205698
107 },
108 {
109 "pair": [
110 0,
111 2
112 ],
113 "names": [
114 "reference",
115 "same_content_diff_sink"
116 ],
117 "cosine": 0.02561184830580232,
118 "js": 0.0590366671190273,
119 "content": 0.0,
120 "sink": 2.1990004360373128,
121 "full": 3.6650007267288562
122 },
123 {
124 "pair": [
125 0,
126 3
127 ],
128 "names": [
129 "reference",
130 "duplicate"
131 ],
132 "cosine": 3.513375346209102e-08,
133 "js": 4.4224831758637043e-07,
134 "content": 4.000326693113483e-05,
135 "sink": 0.0,
136 "full": 0.0002075940850212296
137 },
138 {
139 "pair": [
140 0,
141 4
142 ],
143 "names": [
144 "reference",
145 "unrelated"
146 ],
147 "cosine": 0.14727162513505654,
148 "js": 0.18189119121777383,
149 "content": 5.20819598644114,
150 "sink": 4.334860451017351,
151 "full": 33.265747350567956
152 },
153 {
154 "pair": [
155 1,
156 2
157 ],
158 "names": [
159 "same_sink_diff_content",
160 "same_content_diff_sink"
161 ],
162 "cosine": 0.03251935297043329,
163 "js": 0.09458794154276691,
164 "content": 5.20819598644114,
165 "sink": 2.1990004360373128,
166 "full": 29.705980658934557
167 },
168 {
169 "pair": [
170 1,
171 3
172 ],
173 "names": [
174 "same_sink_diff_content",
175 "duplicate"
176 ],
177 "cosine": 0.0008813247475819574,
178 "js": 0.018583472889377547,
179 "content": 5.21568968913976,
180 "sink": 0.0,
181 "full": 26.078456023449164
182 },
183 {
184 "pair": [
185 1,
186 4
187 ],
188 "names": [
189 "same_sink_diff_content",
190 "unrelated"
191 ],
192 "cosine": 0.13309407222353753,
193 "js": 0.14233756540128167,
194 "content": 0.0,
195 "sink": 4.334860451017351,
196 "full": 7.224767418362257
197 },
198 {
199 "pair": [
200 2,
201 3
202 ],
203 "names": [
204 "same_content_diff_sink",
205 "duplicate"
206 ],
207 "cosine": 0.025588090364458638,
208 "js": 0.05903742478750083,
209 "content": 4.000326693113483e-05,
210 "sink": 2.1990004360373128,
211 "full": 3.675748237435319
212 },
213 {
214 "pair": [
215 2,
216 4
217 ],
218 "names": [
219 "same_content_diff_sink",
220 "unrelated"
221 ],
222 "cosine": 0.1548474175473804,
223 "js": 0.16173372311050377,
224 "content": 5.20819598644114,
225 "sink": 0.35895683210836643,
226 "full": 26.639241319052978
227 },
228 {
229 "pair": [
230 3,
231 4
232 ],
233 "names": [
234 "duplicate",
235 "unrelated"
236 ],
237 "cosine": 0.1472852612992528,
238 "js": 0.18200780679495465,
239 "content": 5.21568968913976,
240 "sink": 4.334860451017351,
241 "full": 33.31802175035605
242 }
243 ],
244 "raw_cosine_nearest": [
245 {
246 "pair": [
247 0,
248 3
249 ],
250 "names": [
251 "reference",
252 "duplicate"
253 ],
254 "cosine": 3.513375346209102e-08,
255 "js": 4.4224831758637043e-07,
256 "content": 4.000326693113483e-05,
257 "sink": 0.0,
258 "full": 0.0002075940850212296
259 },
260 {
261 "pair": [
262 0,
263 1
264 ],
265 "names": [
266 "reference",
267 "same_sink_diff_content"
268 ],
269 "cosine": 0.0008771847295698798,
270 "js": 0.018550749134822735,
271 "content": 5.20819598644114,
272 "sink": 0.0,
273 "full": 26.040979932205698
274 },
275 {
276 "pair": [
277 1,
278 3
279 ],
280 "names": [
281 "same_sink_diff_content",
282 "duplicate"
283 ],
284 "cosine": 0.0008813247475819574,
285 "js": 0.018583472889377547,
286 "content": 5.21568968913976,
287 "sink": 0.0,
288 "full": 26.078456023449164
289 },
290 {
291 "pair": [
292 2,
293 3
294 ],
295 "names": [
296 "same_content_diff_sink",
297 "duplicate"
298 ],
299 "cosine": 0.025588090364458638,
300 "js": 0.05903742478750083,
301 "content": 4.000326693113483e-05,
302 "sink": 2.1990004360373128,
303 "full": 3.675748237435319
304 }
305 ],
306 "channel_admissible": [
307 {
308 "pair": [
309 0,
310 3
311 ],
312 "names": [
313 "reference",
314 "duplicate"
315 ],
316 "cosine": 3.513375346209102e-08,
317 "js": 4.4224831758637043e-07,
318 "content": 4.000326693113483e-05,
319 "sink": 0.0,
320 "full": 0.0002075940850212296
321 }
322 ],
323 "thresholds": {
324 "content": 0.02,
325 "sink": 0.05
326 },
327 "intended_redundant_pair": [
328 0,
329 3
330 ]
331 }
332}