Skip to content

Commit e18f2fc

Browse files
cursoragentFluffyAIcode
cursoragent
and
FluffyAIcode
committed
K3 integrated NIAH gate evidence: arch_correct=1.0 PASS, recall gate FAIL (f_theta v1), identity-restore recall=1.0 (machinery validated)
Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
1 parent 9aa1f51 commit e18f2fc

3 files changed

Lines changed: 952 additions & 0 deletions

File tree

‎results/research/k3_identity_restore_ctx70.json‎

Lines changed: 272 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,272 @@
1+
{
2+
"schema_version": 2,
3+
"kind": "k3_integrated_niah_acceptance",
4+
"config": {
5+
"verifier_id": "google/gemma-4-26B-A4B-it",
6+
"drafter_id": "models/dflash-kakeya-baseline",
7+
"f_theta_dir": "results/research/f_theta_v1",
8+
"f_theta_config": {
9+
"drafter_num_layers": 5,
10+
"drafter_num_kv_heads": 8,
11+
"drafter_head_dim": 128,
12+
"verifier_num_layers": 30,
13+
"verifier_num_kv_heads": 8,
14+
"verifier_head_dim": 256,
15+
"rank": 256,
16+
"verifier_layer_kv_heads": [
17+
8,
18+
8,
19+
8,
20+
8,
21+
8,
22+
2,
23+
8,
24+
8,
25+
8,
26+
8,
27+
8,
28+
2,
29+
8,
30+
8,
31+
8,
32+
8,
33+
8,
34+
2,
35+
8,
36+
8,
37+
8,
38+
8,
39+
8,
40+
2,
41+
8,
42+
8,
43+
8,
44+
8,
45+
8,
46+
2
47+
],
48+
"verifier_layer_head_dims": [
49+
256,
50+
256,
51+
256,
52+
256,
53+
256,
54+
512,
55+
256,
56+
256,
57+
256,
58+
256,
59+
256,
60+
512,
61+
256,
62+
256,
63+
256,
64+
256,
65+
256,
66+
512,
67+
256,
68+
256,
69+
256,
70+
256,
71+
256,
72+
512,
73+
256,
74+
256,
75+
256,
76+
256,
77+
256,
78+
512
79+
]
80+
},
81+
"n_samples": 10,
82+
"sink_size": 4,
83+
"window_size": 64,
84+
"haystack_min_lines": 60,
85+
"haystack_max_lines": 81,
86+
"max_new_tokens": 24,
87+
"seed": 42,
88+
"skip_oracle": true,
89+
"identity_restore": true,
90+
"prompt_token_lens": [
91+
1637,
92+
1378,
93+
1299,
94+
1597,
95+
1278,
96+
1617,
97+
1498,
98+
1238,
99+
1498,
100+
1358
101+
]
102+
},
103+
"results": {
104+
"k3_cross_model": {
105+
"name": "k3_cross_model",
106+
"samples_total": 10,
107+
"samples_correct": 10,
108+
"recall": 1.0,
109+
"mean_latency_s": 5.108704237313941,
110+
"median_latency_s": 5.052127701987047,
111+
"per_sample_decoded": [
112+
"BETA-1409\n//thought\nBETA-1409",
113+
"DELTA-3286\n\nthought\nDELTA-3286\n",
114+
"ORCHID-9935\n10. **Final Answer:** ORCHID-99",
115+
"ETA-1520\n\n\nthought\nthought\nETA-15",
116+
"DELTA-4811\n\n\u66f8\u304d\u8fbc\u307f\u5185\u5bb9\u306b\u57fa\u3065\u304f\u3068\u3001\u79d8\u5bc6\u306e\u30b3\u30fc\u30c9\u306f **DELTA-4",
117+
"ALPHA-4257\n\n \n\n\n \n\n\n \n\n\n \n\n\n \n\n\n \n\n\n \n\n\n",
118+
"DELTA-8359\nnormative_thought\nDELTA-8359",
119+
"MAPLE-3615\n\u092c\u52a9\u3051\u304c\u5fc5\u8981\u306a\u5834\u5408\u306f\u3001\u3044\u3064\u3067\u3082\u304a\u77e5\u3089\u305b\u304f\u3060\u3055\u3044\u3002\n",
120+
"ZETA-5552\n\n\n",
121+
"MAPLE-6514\n\ninclude_thought\nMAPLE-6514"
122+
],
123+
"per_sample_correct": [
124+
true,
125+
true,
126+
true,
127+
true,
128+
true,
129+
true,
130+
true,
131+
true,
132+
true,
133+
true
134+
],
135+
"per_sample_decode_tokens": [
136+
24,
137+
24,
138+
24,
139+
24,
140+
24,
141+
24,
142+
24,
143+
24,
144+
16,
145+
24
146+
],
147+
"per_sample_throughput_tokens_per_sec": [
148+
3.8454941553293023,
149+
4.737514101160626,
150+
4.94522928603496,
151+
4.179844858322699,
152+
4.995243891981496,
153+
4.171802989047664,
154+
4.451344750865639,
155+
5.142973584540285,
156+
4.53123556298915,
157+
4.763504484631587
158+
],
159+
"mean_throughput_tokens_per_sec": 4.576418766490341,
160+
"median_throughput_tokens_per_sec": 4.634374832074888,
161+
"min_throughput_tokens_per_sec": 3.8454941553293023,
162+
"max_throughput_tokens_per_sec": 5.142973584540285
163+
}
164+
},
165+
"attention_window": {
166+
"per_config": {
167+
"k3_cross_model": {
168+
"config": "v04_dlm_restored",
169+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)",
170+
"samples_total": 10,
171+
"effective_keys_at_last_query_mean": 1439.8,
172+
"effective_keys_at_last_query_min": 1238,
173+
"effective_keys_at_last_query_max": 1637,
174+
"effective_keys_at_last_query_median": 1438.0,
175+
"effective_attention_fraction_mean": 1.0,
176+
"effective_attention_fraction_min": 1.0,
177+
"effective_attention_fraction_max": 1.0,
178+
"effective_attention_fraction_median": 1.0,
179+
"per_sample": [
180+
{
181+
"config": "v04_dlm_restored",
182+
"seq_len": 1637,
183+
"effective_keys_at_last_query": 1637,
184+
"effective_attention_fraction": 1.0,
185+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
186+
},
187+
{
188+
"config": "v04_dlm_restored",
189+
"seq_len": 1378,
190+
"effective_keys_at_last_query": 1378,
191+
"effective_attention_fraction": 1.0,
192+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
193+
},
194+
{
195+
"config": "v04_dlm_restored",
196+
"seq_len": 1299,
197+
"effective_keys_at_last_query": 1299,
198+
"effective_attention_fraction": 1.0,
199+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
200+
},
201+
{
202+
"config": "v04_dlm_restored",
203+
"seq_len": 1597,
204+
"effective_keys_at_last_query": 1597,
205+
"effective_attention_fraction": 1.0,
206+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
207+
},
208+
{
209+
"config": "v04_dlm_restored",
210+
"seq_len": 1278,
211+
"effective_keys_at_last_query": 1278,
212+
"effective_attention_fraction": 1.0,
213+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
214+
},
215+
{
216+
"config": "v04_dlm_restored",
217+
"seq_len": 1617,
218+
"effective_keys_at_last_query": 1617,
219+
"effective_attention_fraction": 1.0,
220+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
221+
},
222+
{
223+
"config": "v04_dlm_restored",
224+
"seq_len": 1498,
225+
"effective_keys_at_last_query": 1498,
226+
"effective_attention_fraction": 1.0,
227+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
228+
},
229+
{
230+
"config": "v04_dlm_restored",
231+
"seq_len": 1238,
232+
"effective_keys_at_last_query": 1238,
233+
"effective_attention_fraction": 1.0,
234+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
235+
},
236+
{
237+
"config": "v04_dlm_restored",
238+
"seq_len": 1498,
239+
"effective_keys_at_last_query": 1498,
240+
"effective_attention_fraction": 1.0,
241+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
242+
},
243+
{
244+
"config": "v04_dlm_restored",
245+
"seq_len": 1358,
246+
"effective_keys_at_last_query": 1358,
247+
"effective_attention_fraction": 1.0,
248+
"structural_constraint": "causal_with_dlm_reconstruction (local_cache=sink=4+window=64)"
249+
}
250+
]
251+
}
252+
}
253+
},
254+
"memory": {
255+
"k3_cross_model": {
256+
"device_kind": "cuda",
257+
"device_name": "NVIDIA H200",
258+
"device_total_bytes": 150109880320,
259+
"current_allocated_bytes": 52630830080,
260+
"current_reserved_bytes": 56555995136,
261+
"peak_allocated_bytes": 55607588864,
262+
"peak_reserved_bytes": 56555995136
263+
}
264+
},
265+
"gate": {
266+
"architectural_correctness": true,
267+
"recall_cross_model": 1.0,
268+
"recall_oracle": null,
269+
"recall_delta_vs_oracle_pp": null,
270+
"recall_delta_within_5pp": false
271+
}
272+
}

0 commit comments

Comments
 (0)