johnsonchromia commited on
Commit
bb2a1cb
·
verified ·
1 Parent(s): 9205cf6

Add four Qwen Ambassador benchmark results

Browse files

Publish verified benchmark results and accurately report incomplete runs. Preserve dataset files, original benchmark results and license wording.

Files changed (3) hide show
  1. README.md +6 -2
  2. SHA256SUMS +2 -2
  3. metadata/model-benchmarks-latest.json +347 -3
README.md CHANGED
@@ -64,15 +64,19 @@ The current test has 381 examples, including 70 disagreements, from `outcomes_20
64
  | Kimi K3 | Current, test | 47.83% | 31.43% | 0.4485 |
65
  | GPT-6 Luna | Current, test | 41.00% | 58.57% | 0.3238 |
66
  | GPT-6.1 Sol | Current, test | 60.44% | 78.57% | 0.6037 |
 
 
 
 
67
 
68
  Sol completed the test. Validation was deferred after gateway errors, with saved but unscored predictions: 140/1349 validation.
69
 
70
- The current runs requested low reasoning and a 4,096-token output limit. Sol and Luna ran through a local gateway; the GLMs and Kimi used OpenRouter. Providers and quantization differ, so this compares the recorded serving setups. The gateway did not report dollar costs. [Full results, usage and settings](blob/main/metadata/model-benchmarks-latest.json).
71
 
72
  <details>
73
  <summary>Current validation results</summary>
74
 
75
- There are 1,349 validation examples but only six disagreements. One case changes recall by 16.7 percentage points. Keep validation and test scores separate: disagreement prevalence is 0.44% and 18.37%, respectively.
76
 
77
  | Model | Dataset used | Precision | Recall | Average precision |
78
  | --- | --- | ---: | ---: | ---: |
 
64
  | Kimi K3 | Current, test | 47.83% | 31.43% | 0.4485 |
65
  | GPT-6 Luna | Current, test | 41.00% | 58.57% | 0.3238 |
66
  | GPT-6.1 Sol | Current, test | 60.44% | 78.57% | 0.6037 |
67
+ | Qwen3.8-27B | Current, test | 56.06% | 52.86% | 0.4420 |
68
+ | Qwen3.8-Omni-Flash | Current, test | 29.89% | 37.14% | 0.2677 |
69
+ | Qwen3.8-Flash-Next | Current, test | 40.54% | 42.86% | 0.3006 |
70
+ | Qwen3.8-Max | Current, test | 43.80% | 75.71% | 0.3919 |
71
 
72
  Sol completed the test. Validation was deferred after gateway errors, with saved but unscored predictions: 140/1349 validation.
73
 
74
+ All current test runs use a 4,096-token output limit. The four Qwen Ambassador models ran through ModelScope with thinking disabled; Omni-Flash received text only. The earlier models requested low reasoning: Sol and Luna used a local gateway, while the GLMs and Kimi used OpenRouter. Reasoning settings, providers and quantization differ, so this compares the recorded serving setups. ModelScope and the local gateway did not report dollar costs. [Full results, usage and settings](blob/main/metadata/model-benchmarks-latest.json).
75
 
76
  <details>
77
  <summary>Current validation results</summary>
78
 
79
+ The four Qwen Ambassador models were tested on the current test split only. There are 1,349 validation examples but only six disagreements. One case changes recall by 16.7 percentage points. Keep validation and test scores separate: disagreement prevalence is 0.44% and 18.37%, respectively.
80
 
81
  | Model | Dataset used | Precision | Recall | Average precision |
82
  | --- | --- | ---: | ---: | ---: |
SHA256SUMS CHANGED
@@ -1,10 +1,10 @@
1
  9e75dd981de037ec3769f24f790e126bc5a160b6871f510214e68dc70649aeeb .gitattributes
2
- 21cbe4274da39e5547c466c0f4cc1968ca3946b04ab642320b8c458828f5e309 README.md
3
  cbb49b56ea3c650ac26bcb42261601cc70d36c94689f21047aee89cafd51a86f data/test.jsonl
4
  776042127eb5332b654c8bf3c73a5b182ea8f86aa77d2211bcecd5dd0a6561e7 data/train.jsonl
5
  c430b3d8c59691f64573a8e4ac8db998bb6c9c5c917651814322c720fca8be37 data/validation.jsonl
6
  e8906042012cd72a5f933c4aa63a318bd298e5669f170a85f066401a46f6917f metadata/dataset-card.json
7
- 06e9411199453de7cf012710e0c1f31c4f3551a2374a55e95167ea812687a1f3 metadata/model-benchmarks-latest.json
8
  38c780ea5ab231a1356bd81461e055b3e264fca89b8c901a5665cc45034ad17c metadata/model-benchmarks.json
9
  253447489b29980dbacdb8960b778d33ab10236c50873b93ccd11b53a9b5389e metadata/oracle-audit.json
10
  9d02a7e7c586339aa2618e40493ec72c569545cc2d65806687ca399ba2d8b0e3 metadata/refresh-report.json
 
1
  9e75dd981de037ec3769f24f790e126bc5a160b6871f510214e68dc70649aeeb .gitattributes
2
+ 462d5f702fb4c344eb89b4dfabb236c00ccc0250204020874800b77b367288fb README.md
3
  cbb49b56ea3c650ac26bcb42261601cc70d36c94689f21047aee89cafd51a86f data/test.jsonl
4
  776042127eb5332b654c8bf3c73a5b182ea8f86aa77d2211bcecd5dd0a6561e7 data/train.jsonl
5
  c430b3d8c59691f64573a8e4ac8db998bb6c9c5c917651814322c720fca8be37 data/validation.jsonl
6
  e8906042012cd72a5f933c4aa63a318bd298e5669f170a85f066401a46f6917f metadata/dataset-card.json
7
+ 7aa3f043b95ce5cc82ad701f943e47edaba4d01bc887b59cf4ed706ffce8cb4f metadata/model-benchmarks-latest.json
8
  38c780ea5ab231a1356bd81461e055b3e264fca89b8c901a5665cc45034ad17c metadata/model-benchmarks.json
9
  253447489b29980dbacdb8960b778d33ab10236c50873b93ccd11b53a9b5389e metadata/oracle-audit.json
10
  9d02a7e7c586339aa2618e40493ec72c569545cc2d65806687ca399ba2d8b0e3 metadata/refresh-report.json
metadata/model-benchmarks-latest.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
- "created_at": "2026-10-13T15:05:33.728673+00:00",
3
  "dataset_configuration": "outcomes_2026_10_13",
4
- "dataset_revision": "dbd94f08bce6e52c44b9b17a4bdba133cc96a148",
5
  "task": "attested_vs_disagreed",
6
  "positive_class": "disagreed",
7
  "threshold": 0.5,
@@ -19,7 +19,11 @@
19
  "Recorded cost includes logged failed-output attempts. Transport failures may omit billing information.",
20
  "Sol was added as a later baseline with unchanged prompt, threshold and splits.",
21
  "The local gateway accepted the same request format; upstream checkpoint and effective sampling settings are not independently verified.",
22
- "GLM 5.3 uses Morph FP8 and reported no reasoning tokens. Kimi uses InferenceNet FP4. The abandoned Sail attempt is excluded from scores."
 
 
 
 
23
  ],
24
  "fine_tuning_performed": false,
25
  "test_set_evaluated": true,
@@ -1112,6 +1116,346 @@
1112
  "error": null
1113
  }
1114
  ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1115
  }
1116
  ],
1117
  "incomplete_runs": [
 
1
  {
2
+ "created_at": "2026-10-15T06:20:14.592340+00:00",
3
  "dataset_configuration": "outcomes_2026_10_13",
4
+ "dataset_revision": "9205cf6717f770d57867177ac42fdb233083e645",
5
  "task": "attested_vs_disagreed",
6
  "positive_class": "disagreed",
7
  "threshold": 0.5,
 
19
  "Recorded cost includes logged failed-output attempts. Transport failures may omit billing information.",
20
  "Sol was added as a later baseline with unchanged prompt, threshold and splits.",
21
  "The local gateway accepted the same request format; upstream checkpoint and effective sampling settings are not independently verified.",
22
+ "GLM 5.3 uses Morph FP8 and reported no reasoning tokens. Kimi uses InferenceNet FP4. The abandoned Sail attempt is excluded from scores.",
23
+ "Four Qwen Ambassador baselines were added on the frozen 381-example test only; validation was not run for these models.",
24
+ "Qwen requests use ModelScope, enable_thinking=false, temperature=0, JSON-object output and max_tokens=4096. Mode was selected using synthetic connection checks before benchmark scoring.",
25
+ "Omni-Flash receives text only. ModelScope does not report dollar cost or a reasoning-token breakdown; null values mean unreported, not zero.",
26
+ "Qwen non-thinking runs and earlier low-reasoning runs use different compute settings. Served checkpoint identity is provider-reported, not independently verified."
27
  ],
28
  "fine_tuning_performed": false,
29
  "test_set_evaluated": true,
 
1116
  "error": null
1117
  }
1118
  ]
1119
+ },
1120
+ {
1121
+ "model": "Qwen-Ambassador/Qwen3.8-27B",
1122
+ "split": "test",
1123
+ "completed": 381,
1124
+ "positive_count": 70,
1125
+ "evaluation_sha256": "330b2370c7b4cb3251253e7f1522c35c866b82d5eb5703197636a05d2a89f4b9",
1126
+ "metrics": {
1127
+ "n": 381,
1128
+ "positive_count": 70,
1129
+ "disagreement_prevalence": 0.1837270341207349,
1130
+ "accuracy": 0.8372703412073491,
1131
+ "precision_disagreed": 0.5606060606060606,
1132
+ "recall_disagreed": 0.5285714285714286,
1133
+ "f1_disagreed": 0.5441176470588235,
1134
+ "balanced_accuracy": 0.7176619200734957,
1135
+ "average_precision_disagreed": 0.44199247041339806,
1136
+ "brier": 0.13831364829396342,
1137
+ "confusion": {
1138
+ "tp": 37,
1139
+ "tn": 282,
1140
+ "fp": 29,
1141
+ "fn": 33
1142
+ },
1143
+ "threshold": 0.5,
1144
+ "warning": null
1145
+ },
1146
+ "always_attested": {
1147
+ "n": 381,
1148
+ "positive_count": 70,
1149
+ "disagreement_prevalence": 0.1837270341207349,
1150
+ "accuracy": 0.8162729658792651,
1151
+ "precision_disagreed": 0.0,
1152
+ "recall_disagreed": 0.0,
1153
+ "f1_disagreed": 0.0,
1154
+ "balanced_accuracy": 0.5,
1155
+ "average_precision_disagreed": 0.1837270341207349,
1156
+ "brier": 0.1837270341207349,
1157
+ "confusion": {
1158
+ "tp": 0,
1159
+ "tn": 311,
1160
+ "fp": 0,
1161
+ "fn": 70
1162
+ },
1163
+ "threshold": 0.5,
1164
+ "warning": null
1165
+ },
1166
+ "usage_including_recorded_failed_attempts": {
1167
+ "requests_with_usage": 381,
1168
+ "requests_with_cost": 0,
1169
+ "reported_cost_usd": null,
1170
+ "prompt_tokens": 207660,
1171
+ "completion_tokens": 4809,
1172
+ "requests_with_reasoning_token_breakdown": 0,
1173
+ "reasoning_tokens": null
1174
+ },
1175
+ "failed_attempts": 39,
1176
+ "http_statuses": {
1177
+ "429": 39
1178
+ },
1179
+ "finish_reasons_failed": {},
1180
+ "response_models": {
1181
+ "Qwen-Ambassador/Qwen3.8-27B": 381
1182
+ },
1183
+ "providers": {
1184
+ "None": 381
1185
+ },
1186
+ "settings": {
1187
+ "max_tokens": 4096,
1188
+ "concurrency": 2,
1189
+ "format_retries": 2,
1190
+ "request_delay": 0.75,
1191
+ "rate_limit_retries": 3
1192
+ },
1193
+ "parameters": {
1194
+ "response_format": {
1195
+ "type": "json_object"
1196
+ },
1197
+ "enable_thinking": false
1198
+ },
1199
+ "mean_success_latency_seconds": 1.2635767512388483,
1200
+ "serving": "ModelScope API",
1201
+ "endpoint": "https://api-inference.modelscope.ai/v1/chat/completions",
1202
+ "prediction_file_sha256": "0db2a856e6f72a197f010d697012c86f21c2e3346fc83649911336f5b1575158",
1203
+ "recovery_settings": []
1204
+ },
1205
+ {
1206
+ "model": "Qwen-Ambassador/Qwen3.8-Omni-Flash",
1207
+ "split": "test",
1208
+ "completed": 381,
1209
+ "positive_count": 70,
1210
+ "evaluation_sha256": "330b2370c7b4cb3251253e7f1522c35c866b82d5eb5703197636a05d2a89f4b9",
1211
+ "metrics": {
1212
+ "n": 381,
1213
+ "positive_count": 70,
1214
+ "disagreement_prevalence": 0.1837270341207349,
1215
+ "accuracy": 0.7244094488188977,
1216
+ "precision_disagreed": 0.2988505747126437,
1217
+ "recall_disagreed": 0.37142857142857144,
1218
+ "f1_disagreed": 0.33121019108280253,
1219
+ "balanced_accuracy": 0.5876435461644465,
1220
+ "average_precision_disagreed": 0.26770837829395,
1221
+ "brier": 0.23767907844577038,
1222
+ "confusion": {
1223
+ "tp": 26,
1224
+ "tn": 250,
1225
+ "fp": 61,
1226
+ "fn": 44
1227
+ },
1228
+ "threshold": 0.5,
1229
+ "warning": null
1230
+ },
1231
+ "always_attested": {
1232
+ "n": 381,
1233
+ "positive_count": 70,
1234
+ "disagreement_prevalence": 0.1837270341207349,
1235
+ "accuracy": 0.8162729658792651,
1236
+ "precision_disagreed": 0.0,
1237
+ "recall_disagreed": 0.0,
1238
+ "f1_disagreed": 0.0,
1239
+ "balanced_accuracy": 0.5,
1240
+ "average_precision_disagreed": 0.1837270341207349,
1241
+ "brier": 0.1837270341207349,
1242
+ "confusion": {
1243
+ "tp": 0,
1244
+ "tn": 311,
1245
+ "fp": 0,
1246
+ "fn": 70
1247
+ },
1248
+ "threshold": 0.5,
1249
+ "warning": null
1250
+ },
1251
+ "usage_including_recorded_failed_attempts": {
1252
+ "requests_with_usage": 381,
1253
+ "requests_with_cost": 0,
1254
+ "reported_cost_usd": null,
1255
+ "prompt_tokens": 207660,
1256
+ "completion_tokens": 5488,
1257
+ "requests_with_reasoning_token_breakdown": 0,
1258
+ "reasoning_tokens": null
1259
+ },
1260
+ "failed_attempts": 36,
1261
+ "http_statuses": {
1262
+ "429": 36
1263
+ },
1264
+ "finish_reasons_failed": {},
1265
+ "response_models": {
1266
+ "Qwen-Ambassador/Qwen3.8-Omni-Flash": 381
1267
+ },
1268
+ "providers": {
1269
+ "None": 381
1270
+ },
1271
+ "settings": {
1272
+ "max_tokens": 4096,
1273
+ "concurrency": 2,
1274
+ "format_retries": 2,
1275
+ "request_delay": 0.75,
1276
+ "rate_limit_retries": 3
1277
+ },
1278
+ "parameters": {
1279
+ "response_format": {
1280
+ "type": "json_object"
1281
+ },
1282
+ "enable_thinking": false
1283
+ },
1284
+ "mean_success_latency_seconds": 1.1554911328792676,
1285
+ "serving": "ModelScope API",
1286
+ "endpoint": "https://api-inference.modelscope.ai/v1/chat/completions",
1287
+ "prediction_file_sha256": "1e67af7af2b4fcd8b538f94262a1dde506ee38057add15ef4948a08b8c07b7a0",
1288
+ "recovery_settings": []
1289
+ },
1290
+ {
1291
+ "model": "Qwen-Ambassador/Qwen3.8-Flash-Next",
1292
+ "split": "test",
1293
+ "completed": 381,
1294
+ "positive_count": 70,
1295
+ "evaluation_sha256": "330b2370c7b4cb3251253e7f1522c35c866b82d5eb5703197636a05d2a89f4b9",
1296
+ "metrics": {
1297
+ "n": 381,
1298
+ "positive_count": 70,
1299
+ "disagreement_prevalence": 0.1837270341207349,
1300
+ "accuracy": 0.7795275590551181,
1301
+ "precision_disagreed": 0.40540540540540543,
1302
+ "recall_disagreed": 0.42857142857142855,
1303
+ "f1_disagreed": 0.4166666666666667,
1304
+ "balanced_accuracy": 0.6435461644464859,
1305
+ "average_precision_disagreed": 0.3005844414457841,
1306
+ "brier": 0.18175931758530156,
1307
+ "confusion": {
1308
+ "tp": 30,
1309
+ "tn": 267,
1310
+ "fp": 44,
1311
+ "fn": 40
1312
+ },
1313
+ "threshold": 0.5,
1314
+ "warning": null
1315
+ },
1316
+ "always_attested": {
1317
+ "n": 381,
1318
+ "positive_count": 70,
1319
+ "disagreement_prevalence": 0.1837270341207349,
1320
+ "accuracy": 0.8162729658792651,
1321
+ "precision_disagreed": 0.0,
1322
+ "recall_disagreed": 0.0,
1323
+ "f1_disagreed": 0.0,
1324
+ "balanced_accuracy": 0.5,
1325
+ "average_precision_disagreed": 0.1837270341207349,
1326
+ "brier": 0.1837270341207349,
1327
+ "confusion": {
1328
+ "tp": 0,
1329
+ "tn": 311,
1330
+ "fp": 0,
1331
+ "fn": 70
1332
+ },
1333
+ "threshold": 0.5,
1334
+ "warning": null
1335
+ },
1336
+ "usage_including_recorded_failed_attempts": {
1337
+ "requests_with_usage": 381,
1338
+ "requests_with_cost": 0,
1339
+ "reported_cost_usd": null,
1340
+ "prompt_tokens": 207660,
1341
+ "completion_tokens": 4734,
1342
+ "requests_with_reasoning_token_breakdown": 0,
1343
+ "reasoning_tokens": null
1344
+ },
1345
+ "failed_attempts": 35,
1346
+ "http_statuses": {
1347
+ "429": 35
1348
+ },
1349
+ "finish_reasons_failed": {},
1350
+ "response_models": {
1351
+ "Qwen-Ambassador/Qwen3.8-Flash-Next": 381
1352
+ },
1353
+ "providers": {
1354
+ "None": 381
1355
+ },
1356
+ "settings": {
1357
+ "max_tokens": 4096,
1358
+ "concurrency": 2,
1359
+ "format_retries": 2,
1360
+ "request_delay": 0.75,
1361
+ "rate_limit_retries": 3
1362
+ },
1363
+ "parameters": {
1364
+ "response_format": {
1365
+ "type": "json_object"
1366
+ },
1367
+ "enable_thinking": false
1368
+ },
1369
+ "mean_success_latency_seconds": 1.3607868444960642,
1370
+ "serving": "ModelScope API",
1371
+ "endpoint": "https://api-inference.modelscope.ai/v1/chat/completions",
1372
+ "prediction_file_sha256": "790070ea2d3e0ab21874fcd784f3e5645a2e73d953778b191113ebc0cfe9f0a5",
1373
+ "recovery_settings": []
1374
+ },
1375
+ {
1376
+ "model": "Qwen-Ambassador/Qwen3.8-Max",
1377
+ "split": "test",
1378
+ "completed": 381,
1379
+ "positive_count": 70,
1380
+ "evaluation_sha256": "330b2370c7b4cb3251253e7f1522c35c866b82d5eb5703197636a05d2a89f4b9",
1381
+ "metrics": {
1382
+ "n": 381,
1383
+ "positive_count": 70,
1384
+ "disagreement_prevalence": 0.1837270341207349,
1385
+ "accuracy": 0.7769028871391076,
1386
+ "precision_disagreed": 0.4380165289256198,
1387
+ "recall_disagreed": 0.7571428571428571,
1388
+ "f1_disagreed": 0.5549738219895288,
1389
+ "balanced_accuracy": 0.7692466697289848,
1390
+ "average_precision_disagreed": 0.391917194001231,
1391
+ "brier": 0.17519527559055192,
1392
+ "confusion": {
1393
+ "tp": 53,
1394
+ "tn": 243,
1395
+ "fp": 68,
1396
+ "fn": 17
1397
+ },
1398
+ "threshold": 0.5,
1399
+ "warning": null
1400
+ },
1401
+ "always_attested": {
1402
+ "n": 381,
1403
+ "positive_count": 70,
1404
+ "disagreement_prevalence": 0.1837270341207349,
1405
+ "accuracy": 0.8162729658792651,
1406
+ "precision_disagreed": 0.0,
1407
+ "recall_disagreed": 0.0,
1408
+ "f1_disagreed": 0.0,
1409
+ "balanced_accuracy": 0.5,
1410
+ "average_precision_disagreed": 0.1837270341207349,
1411
+ "brier": 0.1837270341207349,
1412
+ "confusion": {
1413
+ "tp": 0,
1414
+ "tn": 311,
1415
+ "fp": 0,
1416
+ "fn": 70
1417
+ },
1418
+ "threshold": 0.5,
1419
+ "warning": null
1420
+ },
1421
+ "usage_including_recorded_failed_attempts": {
1422
+ "requests_with_usage": 381,
1423
+ "requests_with_cost": 0,
1424
+ "reported_cost_usd": null,
1425
+ "prompt_tokens": 207660,
1426
+ "completion_tokens": 4256,
1427
+ "requests_with_reasoning_token_breakdown": 0,
1428
+ "reasoning_tokens": null
1429
+ },
1430
+ "failed_attempts": 41,
1431
+ "http_statuses": {
1432
+ "429": 41
1433
+ },
1434
+ "finish_reasons_failed": {},
1435
+ "response_models": {
1436
+ "Qwen-Ambassador/Qwen3.8-Max": 381
1437
+ },
1438
+ "providers": {
1439
+ "None": 381
1440
+ },
1441
+ "settings": {
1442
+ "max_tokens": 4096,
1443
+ "concurrency": 2,
1444
+ "format_retries": 2,
1445
+ "request_delay": 0.75,
1446
+ "rate_limit_retries": 3
1447
+ },
1448
+ "parameters": {
1449
+ "response_format": {
1450
+ "type": "json_object"
1451
+ },
1452
+ "enable_thinking": false
1453
+ },
1454
+ "mean_success_latency_seconds": 1.346567590333332,
1455
+ "serving": "ModelScope API",
1456
+ "endpoint": "https://api-inference.modelscope.ai/v1/chat/completions",
1457
+ "prediction_file_sha256": "5f71d37b9ec606e994c428193474810cf2f528faf48dca99be7279f7d8e5ac39",
1458
+ "recovery_settings": []
1459
  }
1460
  ],
1461
  "incomplete_runs": [