File size: 6,685 Bytes
28cf730 | 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 | {
"schema_version": 1,
"status": "complete_independently_audited",
"model": "New Open-Jev 27B on community-hard-mix-v2",
"optimizer_steps": 37160,
"checkpoint": {
"model": "Qwen/Qwen3.8-27B",
"method": "lora_decision_head",
"checkpoint_sha256": "c49994563c3c4f04a99d9130203c4e526f4ae5086c84deec57698d18cb652e71",
"base_revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
"temperature": 2.5343690298472983,
"code_commit": "7ac6bab261bd97efc3d446152999e94ae432e5b0",
"max_length": 16384
},
"scope": "231 public JevBench tasks only: 72 original, 48 easy, 111 hard. The private and judge tiers are unavailable; this is not the full 534-task benchmark and no full-benchmark composite is computed. ZefanCai Open-Jev uses Qwen LoRA weights and a decision head; it is a different system from the Kotoba and Codiv projects also called OpenJev.",
"benchmark": {
"upstream_commit": "f8ce71361165846101d02ebc83ad44e47ae44fc3",
"input_sha256": "2e20ff5f94dd04d015b3a7b9abd955757b6c5be25b55b97fcd141cb52f88b4aa",
"dataset_sha256": "dc3995d8ae1e2fc8e81ce38431add509eb8bb39b85aadfd0c7c32079382dde51"
},
"metrics": {
"n_planned": 231,
"n_attempted": 231,
"n_scorable": 231,
"n_valid": 231,
"n_correct": 197,
"accuracy": 0.8528138528138528,
"schema_validity": 1.0,
"schema_validity_strict": 1.0,
"n_renormalized": 0,
"operational_success": 1.0,
"brier_mean": 0.2419831167860275,
"ordinal_mae": 0.3798003972534468
},
"per_public_tier": {
"original": {
"n_planned": 72,
"n_attempted": 72,
"n_valid": 72,
"n_correct": 69,
"accuracy": 0.9583333333333334
},
"easy": {
"n_planned": 48,
"n_attempted": 48,
"n_valid": 48,
"n_correct": 48,
"accuracy": 1.0
},
"hard": {
"n_planned": 111,
"n_attempted": 111,
"n_valid": 111,
"n_correct": 80,
"accuracy": 0.7207207207207207
}
},
"per_family": {
"policy": {
"n_planned": 12,
"n_correct": 11,
"accuracy": 0.9166666666666666
},
"intent": {
"n_planned": 24,
"n_correct": 23,
"accuracy": 0.9583333333333334
},
"ordinal": {
"n_planned": 12,
"n_correct": 12,
"accuracy": 1.0
},
"extraction": {
"n_planned": 24,
"n_correct": 23,
"accuracy": 0.9583333333333334
},
"adequacy": {
"n_planned": 12,
"n_correct": 12,
"accuracy": 1.0
},
"routing": {
"n_planned": 12,
"n_correct": 12,
"accuracy": 1.0
},
"fact": {
"n_planned": 12,
"n_correct": 12,
"accuracy": 1.0
},
"tool_selection": {
"n_planned": 12,
"n_correct": 12,
"accuracy": 1.0
},
"long_policy": {
"n_planned": 19,
"n_correct": 11,
"accuracy": 0.5789473684210527
},
"probability": {
"n_planned": 10,
"n_correct": 7,
"accuracy": 0.7
},
"temporal_numeric": {
"n_planned": 15,
"n_correct": 2,
"accuracy": 0.13333333333333333
},
"ambiguous": {
"n_planned": 7,
"n_correct": 5,
"accuracy": 0.7142857142857143
},
"multi_hop": {
"n_planned": 18,
"n_correct": 16,
"accuracy": 0.8888888888888888
},
"tradeoff": {
"n_planned": 6,
"n_correct": 5,
"accuracy": 0.8333333333333334
},
"adversarial": {
"n_planned": 6,
"n_correct": 6,
"accuracy": 1.0
},
"trap": {
"n_planned": 8,
"n_correct": 8,
"accuracy": 1.0
},
"judge_hard": {
"n_planned": 17,
"n_correct": 15,
"accuracy": 0.8823529411764706
},
"routing_hard": {
"n_planned": 5,
"n_correct": 5,
"accuracy": 1.0
}
},
"comparison_baselines": [
{
"label": "Released Open-Jev 2B",
"overall": {
"n_correct": 150,
"n_planned": 231,
"accuracy": 0.6493506493506493
},
"hard": {
"n_correct": 46,
"n_planned": 111,
"accuracy": 0.4144144144144144
}
},
{
"label": "Released Open-Jev 9B",
"overall": {
"n_correct": 179,
"n_planned": 231,
"accuracy": 0.7748917748917749
},
"hard": {
"n_correct": 66,
"n_planned": 111,
"accuracy": 0.5945945945945946
}
},
{
"label": "Jev 1.13.0",
"overall": {
"n_correct": 200,
"n_planned": 231,
"accuracy": 0.8658008658008658
},
"hard": {
"n_correct": 81,
"n_planned": 111,
"accuracy": 0.7297297297297297
}
}
],
"protocol": {
"sum_tolerance_strict": 0.001,
"sum_tolerance_rounding": 0.02,
"argmax_tie_break": "lexicographically_smallest_label",
"score_accuracy": "argmax_level_equals_gold_level",
"probabilities": "native",
"concurrency": 1,
"retries": 0,
"warmups": 0,
"prefix_cache": false,
"latency_note": "Four-rank frozen-base FSDP2 execution on a shared node; one observation per task. Timing starts after dispatch and includes collective prediction, rank validation and response validation. No HTTP/network timing, warmups or retries. These diagnostics are not comparable to the old single-GPU HTTP latency, matched-hardware speedups, throughput or energy efficiency.",
"transport": "torch_distributed_collective",
"execution": {
"mode": "frozen_base_fsdp2",
"world_size": 4,
"backend": "nccl",
"node_shared_with_other_jobs": true,
"candidate_batch_size": 1,
"http_server_started": false
}
},
"execution": {
"mode": "frozen_base_fsdp2",
"world_size": 4,
"backend": "nccl",
"node_shared_with_other_jobs": true,
"candidate_batch_size": 1,
"http_server_started": false
},
"failed_requests": 0,
"pending_requests": 0,
"in_flight_requests": 0,
"audit": {
"file": "audit.json",
"sha256": "0759d6d9d457e0b823f823d8d9e7b7c9f0f4f3a9278f6cd32aac1e7b442392fc"
},
"source_summary_sha256": "ede7fda4e114c22bed02ec5a389a3bd4b9952eeeeed4a64e39cdce88d6589b8f",
"limitations": [
"231 public tasks only; unavailable private/judge tasks were not scored.",
"Four-rank shared-node direct-execution timings cannot be compared as speedups against the old single-GPU HTTP or hosted HTTPS timings.",
"Checkpoint hash verification establishes artifact identity; this offline audit does not rerun model inference or establish benchmark-training semantic separation.",
"Raw requests, labels, task IDs and responses remain in an ignored local run directory."
]
}
|