Xananthium commited on
Commit
eb5ce07
·
verified ·
1 Parent(s): b312337

Publish hauhau-qwen3.6-35b-source with measured results and source credits

Browse files
.gitattributes CHANGED
@@ -35,3 +35,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  qwen36-int8-mixed-probe/tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
  qwen36-int8-sq-export-probe/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
 
 
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  qwen36-int8-mixed-probe/tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
  qwen36-int8-sq-export-probe/tokenizer.json filter=lfs diff=lfs merge=lfs -text
38
+ hauhau-qwen3.6-35b-source/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_0-vLLM.gguf filter=lfs diff=lfs merge=lfs -text
39
+ hauhau-qwen3.6-35b-source/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf filter=lfs diff=lfs merge=lfs -text
40
+ hauhau-qwen3.6-35b-source/mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf filter=lfs diff=lfs merge=lfs -text
hauhau-qwen3.6-35b-source/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_0-vLLM.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:99942372e5a478de1e2312bfa706570c9a677c00ae9b005feee5081a75b9292b
3
+ size 37061900416
hauhau-qwen3.6-35b-source/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5efc2a1fc3f595893128550cf8b7e792774be6c0409c4d2e4729d2b5bb51173
3
+ size 43605014656
hauhau-qwen3.6-35b-source/README.md ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
4
+ tags:
5
+ - local-inference
6
+ - quantization
7
+ - evaluation
8
+ ---
9
+
10
+ # hauhau-qwen3.6-35b-source
11
+
12
+ Complete local checkpoint published by Xananthium, 2026-10-06. Not evaluated in the 2026-10-06 matched precision comparison. The attached report describes other checkpoints, not scores for this artifact.
13
+
14
+ Source credit: [HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive](https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive). Original base: [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B). Original model card is preserved in `README_ORIGINAL.md`; upstream licenses apply.
15
+
16
+ Measured results and limitations are in `results/precision-comparison.json`. These are small synthetic tests, not published cybersecurity benchmark scores. Untested artifacts have no measured scores.
17
+
hauhau-qwen3.6-35b-source/mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c8e702344a81f8c226a914aa980ed6e1f604bce9374f1fed8e65c896908af414
3
+ size 899283072
hauhau-qwen3.6-35b-source/results/precision-comparison.json ADDED
@@ -0,0 +1,382 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "date_utc": "2026-10-06",
3
+ "hardware": "2 x NVIDIA RTX 3090 24 GiB, NVLink, 125 GiB host RAM",
4
+ "vllm": "0.31.0",
5
+ "context_limit": 200000,
6
+ "kv_cache": "turboquant_4bit_nc",
7
+ "host_kv_offload_gib": 90,
8
+ "tensor_parallel_size": 2,
9
+ "sampling": {
10
+ "temperature": 0.6,
11
+ "top_p": 0.95,
12
+ "top_k": 20,
13
+ "presence_penalty": 0.0
14
+ },
15
+ "thinking": true,
16
+ "variants": {
17
+ "hauhau-w8a8": {
18
+ "long_context": {
19
+ "variant": "hauhau-w8a8",
20
+ "baseline": "model-tools/endy-long-context-20261005T172133Z",
21
+ "turns": 8,
22
+ "checks_passed": 53,
23
+ "checks_total": 54,
24
+ "strict_json_turns": 7,
25
+ "tool_calls": 2,
26
+ "median_decode_tokens_per_second": 46.885000000000005,
27
+ "grading_note": "Original checks retained; inspect semantic cyber findings separately."
28
+ },
29
+ "cybersecurity": {
30
+ "variant": "hauhau-w8a8",
31
+ "cases": 6,
32
+ "checks_passed": 30,
33
+ "checks_total": 43,
34
+ "strict_json_cases": 4,
35
+ "checks_passed_after_clarification": 34,
36
+ "note": "Small synthetic assessment, not a published benchmark or comprehensive security validation."
37
+ },
38
+ "speed": {
39
+ "variant": "hauhau-w8a8",
40
+ "output_tokens_per_request": 1024,
41
+ "sampling": {
42
+ "temperature": 0.6,
43
+ "top_p": 0.95,
44
+ "top_k": 20,
45
+ "presence_penalty": 0
46
+ },
47
+ "thinking_enabled": true,
48
+ "requests": [
49
+ {
50
+ "label": "short_0",
51
+ "input_tokens": 34,
52
+ "output_tokens": 1024,
53
+ "wall_seconds": 6.534,
54
+ "first_token_seconds": 0.753,
55
+ "decode_tokens_per_second": 177.144,
56
+ "warmup": true
57
+ },
58
+ {
59
+ "label": "short_1",
60
+ "input_tokens": 34,
61
+ "output_tokens": 1024,
62
+ "wall_seconds": 6.022,
63
+ "first_token_seconds": 0.126,
64
+ "decode_tokens_per_second": 173.698,
65
+ "warmup": false
66
+ },
67
+ {
68
+ "label": "short_2",
69
+ "input_tokens": 34,
70
+ "output_tokens": 1024,
71
+ "wall_seconds": 5.938,
72
+ "first_token_seconds": 0.116,
73
+ "decode_tokens_per_second": 175.89,
74
+ "warmup": false
75
+ },
76
+ {
77
+ "label": "long_0",
78
+ "input_tokens": 121616,
79
+ "output_tokens": 1024,
80
+ "wall_seconds": 40.625,
81
+ "first_token_seconds": 18.331,
82
+ "decode_tokens_per_second": 45.932,
83
+ "warmup": true
84
+ },
85
+ {
86
+ "label": "long_1",
87
+ "input_tokens": 121616,
88
+ "output_tokens": 1024,
89
+ "wall_seconds": 23.206,
90
+ "first_token_seconds": 1.08,
91
+ "decode_tokens_per_second": 46.28,
92
+ "warmup": false
93
+ },
94
+ {
95
+ "label": "long_2",
96
+ "input_tokens": 121616,
97
+ "output_tokens": 1024,
98
+ "wall_seconds": 23.118,
99
+ "first_token_seconds": 1.079,
100
+ "decode_tokens_per_second": 46.462,
101
+ "warmup": false
102
+ }
103
+ ],
104
+ "short_median_decode_tps": 174.79399999999998,
105
+ "long_median_decode_tps": 46.371,
106
+ "four_concurrent": {
107
+ "wall_seconds": 11.758,
108
+ "aggregate_output_tokens_per_second": 348.353,
109
+ "requests": [
110
+ {
111
+ "label": "concurrent_0",
112
+ "input_tokens": 34,
113
+ "output_tokens": 1024,
114
+ "wall_seconds": 11.74,
115
+ "first_token_seconds": 0.128,
116
+ "decode_tokens_per_second": 88.188
117
+ },
118
+ {
119
+ "label": "concurrent_1",
120
+ "input_tokens": 34,
121
+ "output_tokens": 1024,
122
+ "wall_seconds": 11.754,
123
+ "first_token_seconds": 3.292,
124
+ "decode_tokens_per_second": 121.005
125
+ },
126
+ {
127
+ "label": "concurrent_2",
128
+ "input_tokens": 34,
129
+ "output_tokens": 1024,
130
+ "wall_seconds": 11.753,
131
+ "first_token_seconds": 3.291,
132
+ "decode_tokens_per_second": 121.006
133
+ },
134
+ {
135
+ "label": "concurrent_3",
136
+ "input_tokens": 34,
137
+ "output_tokens": 1024,
138
+ "wall_seconds": 11.752,
139
+ "first_token_seconds": 3.29,
140
+ "decode_tokens_per_second": 121.009
141
+ }
142
+ ]
143
+ }
144
+ },
145
+ "custom_metadata": {
146
+ "valid_tool_cases": 6,
147
+ "total_cases": 6,
148
+ "correct_fields": 30,
149
+ "total_fields": 30,
150
+ "raw_json_cases": 5
151
+ },
152
+ "executed_sql_fixture": {
153
+ "passed": 2,
154
+ "total": 2
155
+ },
156
+ "native_api_integration": [
157
+ {
158
+ "test": "model_alias_and_underlying_checkpoint",
159
+ "passed": true
160
+ },
161
+ {
162
+ "test": "anthropic_text_through_caddy",
163
+ "passed": true
164
+ },
165
+ {
166
+ "test": "anthropic_streaming_through_caddy",
167
+ "passed": true
168
+ },
169
+ {
170
+ "test": "anthropic_tool_call_and_result_roundtrip",
171
+ "passed": true
172
+ },
173
+ {
174
+ "test": "anthropic_vision_two_distinct_images",
175
+ "passed": true
176
+ },
177
+ {
178
+ "test": "eight_concurrent_anthropic_requests",
179
+ "passed": true,
180
+ "wall_seconds": 1.0,
181
+ "total_output_tokens": 578
182
+ },
183
+ {
184
+ "test": "public_https_anthropic_endpoint",
185
+ "passed": true
186
+ }
187
+ ],
188
+ "claude_code": {
189
+ "wall_seconds": 2.721,
190
+ "exit_code": 0
191
+ }
192
+ },
193
+ "hauhau-w8a16": {
194
+ "long_context": {
195
+ "variant": "hauhau-w8a16",
196
+ "baseline": "/home/isnotreal/model-tools/endy-long-context-20261005T172133Z",
197
+ "turns": 8,
198
+ "checks_passed": 53,
199
+ "checks_total": 54,
200
+ "strict_json_turns": 7,
201
+ "tool_calls": 2,
202
+ "median_decode_tokens_per_second": 47.269999999999996,
203
+ "grading_note": "Original checks retained; inspect semantic cyber findings separately."
204
+ },
205
+ "cybersecurity": {
206
+ "variant": "hauhau-w8a16",
207
+ "cases": 6,
208
+ "checks_passed": 28,
209
+ "checks_total": 43,
210
+ "strict_json_cases": 6,
211
+ "checks_passed_after_clarification": 36,
212
+ "note": "Small synthetic assessment, not a published benchmark or comprehensive security validation."
213
+ },
214
+ "speed": {
215
+ "variant": "hauhau-w8a16",
216
+ "output_tokens_per_request": 1024,
217
+ "sampling": {
218
+ "temperature": 0.6,
219
+ "top_p": 0.95,
220
+ "top_k": 20,
221
+ "presence_penalty": 0
222
+ },
223
+ "thinking_enabled": true,
224
+ "requests": [
225
+ {
226
+ "label": "short_0",
227
+ "input_tokens": 34,
228
+ "output_tokens": 1024,
229
+ "wall_seconds": 5.891,
230
+ "first_token_seconds": 0.131,
231
+ "decode_tokens_per_second": 177.768,
232
+ "warmup": true
233
+ },
234
+ {
235
+ "label": "short_1",
236
+ "input_tokens": 34,
237
+ "output_tokens": 1024,
238
+ "wall_seconds": 6.107,
239
+ "first_token_seconds": 0.113,
240
+ "decode_tokens_per_second": 170.834,
241
+ "warmup": false
242
+ },
243
+ {
244
+ "label": "short_2",
245
+ "input_tokens": 34,
246
+ "output_tokens": 1024,
247
+ "wall_seconds": 6.021,
248
+ "first_token_seconds": 0.115,
249
+ "decode_tokens_per_second": 173.38,
250
+ "warmup": false
251
+ },
252
+ {
253
+ "label": "long_0",
254
+ "input_tokens": 121616,
255
+ "output_tokens": 1024,
256
+ "wall_seconds": 44.096,
257
+ "first_token_seconds": 21.895,
258
+ "decode_tokens_per_second": 46.124,
259
+ "warmup": true
260
+ },
261
+ {
262
+ "label": "long_1",
263
+ "input_tokens": 121616,
264
+ "output_tokens": 1024,
265
+ "wall_seconds": 23.291,
266
+ "first_token_seconds": 1.18,
267
+ "decode_tokens_per_second": 46.312,
268
+ "warmup": false
269
+ },
270
+ {
271
+ "label": "long_2",
272
+ "input_tokens": 121616,
273
+ "output_tokens": 1024,
274
+ "wall_seconds": 23.12,
275
+ "first_token_seconds": 1.17,
276
+ "decode_tokens_per_second": 46.652,
277
+ "warmup": false
278
+ }
279
+ ],
280
+ "short_median_decode_tps": 172.107,
281
+ "long_median_decode_tps": 46.482,
282
+ "four_concurrent": {
283
+ "wall_seconds": 10.924,
284
+ "aggregate_output_tokens_per_second": 374.959,
285
+ "requests": [
286
+ {
287
+ "label": "concurrent_0",
288
+ "input_tokens": 34,
289
+ "output_tokens": 1024,
290
+ "wall_seconds": 10.907,
291
+ "first_token_seconds": 0.118,
292
+ "decode_tokens_per_second": 94.911
293
+ },
294
+ {
295
+ "label": "concurrent_1",
296
+ "input_tokens": 34,
297
+ "output_tokens": 1024,
298
+ "wall_seconds": 10.914,
299
+ "first_token_seconds": 0.221,
300
+ "decode_tokens_per_second": 95.758
301
+ },
302
+ {
303
+ "label": "concurrent_2",
304
+ "input_tokens": 34,
305
+ "output_tokens": 1024,
306
+ "wall_seconds": 10.921,
307
+ "first_token_seconds": 2.561,
308
+ "decode_tokens_per_second": 122.487
309
+ },
310
+ {
311
+ "label": "concurrent_3",
312
+ "input_tokens": 34,
313
+ "output_tokens": 1024,
314
+ "wall_seconds": 10.92,
315
+ "first_token_seconds": 2.56,
316
+ "decode_tokens_per_second": 122.494
317
+ }
318
+ ]
319
+ }
320
+ },
321
+ "custom_metadata": {
322
+ "valid_tool_cases": 6,
323
+ "total_cases": 6,
324
+ "correct_fields": 30,
325
+ "total_fields": 30,
326
+ "raw_json_cases": 4
327
+ },
328
+ "executed_sql_fixture": {
329
+ "passed": 2,
330
+ "total": 2
331
+ }
332
+ },
333
+ "endy-w8a16": {
334
+ "long_context": {
335
+ "variant": "endy-w8a16-vllm-0.31-qwen-coding-preset",
336
+ "baseline": "/home/isnotreal/model-tools/endy-long-context-20261005T172133Z",
337
+ "turns": 8,
338
+ "checks_passed": 46,
339
+ "checks_total": 54,
340
+ "strict_json_turns": 0,
341
+ "tool_calls": 0,
342
+ "median_decode_tokens_per_second": 47.545,
343
+ "grading_note": "Original checks retained; inspect semantic cyber findings separately."
344
+ },
345
+ "custom_metadata": {
346
+ "fresh_valid_tool_cases": 2,
347
+ "fresh_cases": 2,
348
+ "compacted_valid_tool_cases": 0,
349
+ "compacted_cases": 2,
350
+ "long_valid_tool_cases": 0,
351
+ "long_cases": 2
352
+ },
353
+ "note": "Endy can call tools on fresh prompts. Reliability degraded on the compacted and long-history custom-metadata fixtures."
354
+ },
355
+ "endy-w8a8": {
356
+ "long_context": {
357
+ "variant": "endy-w8a8-vllm-0.31",
358
+ "baseline": "/home/isnotreal/model-tools/endy-long-context-20261005T172133Z",
359
+ "turns": 8,
360
+ "checks_passed": 27,
361
+ "checks_total": 54,
362
+ "strict_json_turns": 0,
363
+ "tool_calls": 0,
364
+ "median_decode_tokens_per_second": 47.515,
365
+ "grading_note": "Original checks retained; inspect semantic cyber findings separately."
366
+ },
367
+ "note": "Earlier W8A8 long-context run used different sampling; not a matched activation-precision comparison."
368
+ }
369
+ },
370
+ "limitations": [
371
+ "Small synthetic local assessment, not a published cybersecurity benchmark.",
372
+ "Long-context tests independently replayed eight original 122k-124k-token prefixes, not an eight-turn fresh conversation.",
373
+ "The one failed long-context strict check per Hauhau variant was a literal regex mismatch; the proposed SQL was executed successfully in a fixture.",
374
+ "The four-concurrent speed measurement is one batch and can include first-shape compilation; it is not a stable concurrency ranking.",
375
+ "Both variants made substantive security-policy mistakes on the broader quiz; manual review is necessary.",
376
+ "W8A8 SSRF clarification exhausted the 4096-output-token cap and truncated JSON; its secondary score includes that failure.",
377
+ "The native W8A8 export uses dynamic INT8 activations throughout eligible language linears; the older SmoothQuant export is a different mixed W8A8/W8A16 artifact.",
378
+ "Hauhau weights originate in Q8_K_P GGUF, then reconstruction and INT8 conversion. W8A16 and W8A8 use identical INT8 weight values, but native kernels and scale handling differ.",
379
+ "MTP and vision weights are retained; speculative decoding was not enabled.",
380
+ "Two serial repeats after warmup and one concurrency batch are too few to attribute small differences to precision alone."
381
+ ]
382
+ }