Publish public and private split benchmark results

README.md

@@ -0,0 +1,87 @@
+---
+language: []
+pretty_name: Perch Benchmark Results
+---
+
+# Perch benchmark results
+
+Bug and security results are separate. The primary tables compare models given the same saved Perch method context and one broad question per method (`perch-broad-v1`). Average precision (AP) measures ranking across all scores; the other metrics use an uncalibrated 0.5 cutoff.
+
+The public pairs and their per-method predictions are available here. Private pair-level data and predictions are not published. The public and private sets are disjoint by source repository across both domains.
+
+**Private-score limitation:** the split was made after these models had been evaluated on the complete collection. The private results below are retrospective, not an untouched blind test. A future blind track needs new pairs that have never been used for model evaluation or selection.
+
+## Bugs
+
+### Public — 2,220 pairs
+
+[Download public bug pairs](https://huggingface.co/datasets/perchscan/bug-pairs-public)
+
+| Model | AP | Precision@0.5 | Recall@0.5 | F1@0.5 | TP | FP | FN |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| Hosted Jev 1.13.0 | 47.6% | 49.9% | 22.0% | 30.6% | 410 | 412 | 1451 |
+| Released JevK5 | 44.1% | 47.5% | 7.6% | 13.1% | 141 | 156 | 1720 |
+| Released Gemma 4 E4B + Jev-lite | 43.7% | 44.8% | 2.3% | 4.4% | 43 | 53 | 1818 |
+
+### Private — 498 pairs
+
+| Model | AP | Precision@0.5 | Recall@0.5 | F1@0.5 | TP | FP | FN |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| Hosted Jev 1.13.0 | 48.6% | 50.7% | 24.1% | 32.6% | 102 | 99 | 322 |
+| Released JevK5 | 45.8% | 46.4% | 9.2% | 15.4% | 39 | 45 | 385 |
+| Released Gemma 4 E4B + Jev-lite | 45.2% | 60.0% | 3.5% | 6.7% | 15 | 10 | 409 |
+
+## Security
+
+### Public — 1,311 pairs
+
+[Download public security pairs](https://huggingface.co/datasets/perchscan/security-pairs-public)
+
+| Model | AP | Precision@0.5 | Recall@0.5 | F1@0.5 | TP | FP | FN |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| Hosted Jev 1.13.0 | 49.2% | 54.4% | 10.2% | 17.2% | 117 | 98 | 1025 |
+| Released JevK5 | 44.8% | 46.4% | 12.4% | 19.6% | 142 | 164 | 1000 |
+| Released Gemma 4 E4B + Jev-lite | 43.3% | 36.0% | 3.2% | 5.8% | 36 | 64 | 1106 |
+
+### Private — 439 pairs
+
+| Model | AP | Precision@0.5 | Recall@0.5 | F1@0.5 | TP | FP | FN |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| Hosted Jev 1.13.0 | 48.9% | 52.6% | 2.5% | 4.8% | 10 | 9 | 388 |
+| Released JevK5 | 47.4% | 45.0% | 4.5% | 8.2% | 18 | 22 | 380 |
+| Released Gemma 4 E4B + Jev-lite | 47.3% | 43.8% | 1.8% | 3.4% | 7 | 9 | 391 |
+
+## Different input protocols
+
+These released models received their native decision inputs, so their scores are reported separately from the prompt-matched tables above.
+
+| Domain | Track | Model | Protocol | AP | F1@0.5 | TP | FP | FN |
+| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
+| Bug | Public | Released CLM-v0.1-8B | `clm-native-choice-6144-v1` | 44.0% | 21.2% | 258 | 317 | 1603 |
+| Bug | Private | Released CLM-v0.1-8B | `clm-native-choice-6144-v1` | 45.2% | 25.1% | 74 | 92 | 350 |
+| Security | Public | Released CLM-v0.1-8B | `clm-native-choice-6144-v1` | 44.0% | 52.1% | 741 | 959 | 401 |
+| Security | Public | Released SecureBERT2.0 | `perch-security-state-4096-v1` | 43.6% | 60.6% | 1140 | 1479 | 2 |
+| Security | Private | Released CLM-v0.1-8B | `clm-native-choice-6144-v1` | 47.1% | 55.4% | 270 | 306 | 128 |
+| Security | Private | Released SecureBERT2.0 | `perch-security-state-4096-v1` | 43.0% | 62.4% | 397 | 478 | 1 |
+
+At 0.5, SecureBERT2.0 flags nearly every security method. Its high F1 therefore does not indicate useful alerting precision.
+
+## Protocol and versions
+
+Each before and after method was scored independently without showing the repair. All reported runs cover every side of their split. `perch-broad-v1` uses one broad bug or security question, not the full multi-question Perch scan. CLM uses its native choice head with a 6,144-token encoder; SecureBERT2.0 uses a 4,096-token input cap. Different protocols are not a prompt-matched ranking. No cutoff was selected on these benchmark labels.
+
+| Model | Frozen checkpoint |
+| --- | --- |
+| Hosted Jev | Service version 1.13.0; provider weights not public |
+| JevK5 | `alibiserikbay/JevK5@3c673298eb7f2dc7cb98019262c55b1fac01a0bc` |
+| Gemma 4 E4B + Jev-lite | Base `ee0ef6023621cff504d758262d4e04895a5af4a2`; released adapter `90b23bc26c53f5a897684456893b2b7edeb90297` |
+| CLM-v0.1-8B | Encoder `b968826d9c46dd6066d109eabc6255188de91218`; released head SHA-256 `b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5` |
+| SecureBERT2.0 | Released `cisco-ai/SecureBERT2.0-code-vuln-detection`; exact snapshot revision was not recorded |
+
+The collection is enriched for before/after changes, so its positive rate and precision do not estimate production-scan prevalence. Some security labels are inferred from advisory fixes rather than independently confirmed for every method.
+
+## Files
+
+- `results.jsonl`: one aggregate report per domain, track, model, and protocol.
+- `split-summary.json`: split sizes and whole-file SHA-256 values. It contains no private rows or repository list.
+- `predictions/`: public-pair probabilities for auditing the public scores. No private predictions are included.

results.jsonl

@@ -0,0 +1,18 @@
+{"average_precision": 0.47645165165489634, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "confusion": {"fn": 1451, "fp": 412, "tn": 2167, "tp": 410}, "domain": "bug", "f1": 0.3056280283265002, "false_positives_per_true_positive": 1.0048780487804878, "model_id": "jev", "model_revision": "hosted-1.13.0", "precision": 0.49878345498783455, "protocol_id": "perch-broad-v1", "recall": 0.22031166039763567, "retrospective_split": true, "run_id": "jev-1.13.0", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 757, "eligible": 1782, "ties": 282}}
+{"average_precision": 0.44111045802808563, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "confusion": {"fn": 1720, "fp": 156, "tn": 2423, "tp": 141}, "domain": "bug", "f1": 0.13067655236329936, "false_positives_per_true_positive": 1.1063829787234043, "model_id": "original-jevk5", "model_revision": "alibiserikbay/JevK5@3c673298eb7f2dc7cb98019262c55b1fac01a0bc", "precision": 0.47474747474747475, "protocol_id": "perch-broad-v1", "recall": 0.07576571735626007, "retrospective_split": true, "run_id": "jevk5-base", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 572, "eligible": 1782, "ties": 463}}
+{"average_precision": 0.43650527821849755, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "confusion": {"fn": 1818, "fp": 53, "tn": 2526, "tp": 43}, "domain": "bug", "f1": 0.04394481349003577, "false_positives_per_true_positive": 1.2325581395348837, "model_id": "vagmi/jev-lite", "model_revision": "gemma-4-E4B-it+jev-lite-released", "precision": 0.4479166666666667, "protocol_id": "perch-broad-v1", "recall": 0.0231058570660935, "retrospective_split": true, "run_id": "gemma-jev-lite-base", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 399, "eligible": 1782, "ties": 844}}
+{"average_precision": 0.43970483059163157, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "confusion": {"fn": 1603, "fp": 317, "tn": 2262, "tp": 258}, "domain": "bug", "f1": 0.21182266009852216, "false_positives_per_true_positive": 1.2286821705426356, "model_id": "Contrastive-LM/CLM-v0.1-8B", "model_revision": "b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5", "precision": 0.44869565217391305, "protocol_id": "clm-native-choice-6144-v1", "recall": 0.138635142396561, "retrospective_split": true, "run_id": "clm-base", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 835, "eligible": 1782, "ties": 1}}
+{"average_precision": 0.4863972179016045, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "confusion": {"fn": 322, "fp": 99, "tn": 473, "tp": 102}, "domain": "bug", "f1": 0.3264, "false_positives_per_true_positive": 0.9705882352941176, "model_id": "jev", "model_revision": "hosted-1.13.0", "precision": 0.5074626865671642, "protocol_id": "perch-broad-v1", "recall": 0.24056603773584906, "retrospective_split": true, "run_id": "jev-1.13.0", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 174, "eligible": 401, "ties": 65}}
+{"average_precision": 0.458256742724197, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "confusion": {"fn": 385, "fp": 45, "tn": 527, "tp": 39}, "domain": "bug", "f1": 0.15354330708661418, "false_positives_per_true_positive": 1.1538461538461537, "model_id": "original-jevk5", "model_revision": "alibiserikbay/JevK5@3c673298eb7f2dc7cb98019262c55b1fac01a0bc", "precision": 0.4642857142857143, "protocol_id": "perch-broad-v1", "recall": 0.09198113207547169, "retrospective_split": true, "run_id": "jevk5-base", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 131, "eligible": 401, "ties": 104}}
+{"average_precision": 0.4523379371851684, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "confusion": {"fn": 409, "fp": 10, "tn": 562, "tp": 15}, "domain": "bug", "f1": 0.066815144766147, "false_positives_per_true_positive": 0.6666666666666666, "model_id": "vagmi/jev-lite", "model_revision": "gemma-4-E4B-it+jev-lite-released", "precision": 0.6, "protocol_id": "perch-broad-v1", "recall": 0.03537735849056604, "retrospective_split": true, "run_id": "gemma-jev-lite-base", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 91, "eligible": 401, "ties": 192}}
+{"average_precision": 0.4521535275641103, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "confusion": {"fn": 350, "fp": 92, "tn": 480, "tp": 74}, "domain": "bug", "f1": 0.25084745762711863, "false_positives_per_true_positive": 1.2432432432432432, "model_id": "Contrastive-LM/CLM-v0.1-8B", "model_revision": "b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5", "precision": 0.4457831325301205, "protocol_id": "clm-native-choice-6144-v1", "recall": 0.17452830188679244, "retrospective_split": true, "run_id": "clm-base", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 174, "eligible": 401, "ties": 0}}
+{"average_precision": 0.49155899716131013, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "confusion": {"fn": 1025, "fp": 98, "tn": 1382, "tp": 117}, "domain": "security", "f1": 0.17243920412675018, "false_positives_per_true_positive": 0.8376068376068376, "model_id": "jev", "model_revision": "hosted-1.13.0", "precision": 0.5441860465116279, "protocol_id": "perch-broad-v1", "recall": 0.10245183887915937, "retrospective_split": true, "run_id": "jev-1.13.0", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 455, "eligible": 1124, "ties": 218}}
+{"average_precision": 0.4476401588790277, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "confusion": {"fn": 1000, "fp": 164, "tn": 1316, "tp": 142}, "domain": "security", "f1": 0.19613259668508287, "false_positives_per_true_positive": 1.1549295774647887, "model_id": "alibiserikbay/JevK5", "model_revision": "3c673298eb7f2dc7cb98019262c55b1fac01a0bc", "precision": 0.46405228758169936, "protocol_id": "perch-broad-v1", "recall": 0.1243432574430823, "retrospective_split": true, "run_id": "jevk5-base", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 344, "eligible": 1124, "ties": 173}}
+{"average_precision": 0.43340784567690765, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "confusion": {"fn": 1106, "fp": 64, "tn": 1416, "tp": 36}, "domain": "security", "f1": 0.057971014492753624, "false_positives_per_true_positive": 1.7777777777777777, "model_id": "vagmi/jev-lite", "model_revision": "gemma-4-E4B-it+jev-lite-released", "precision": 0.36, "protocol_id": "perch-broad-v1", "recall": 0.03152364273204904, "retrospective_split": true, "run_id": "gemma-jev-lite-base", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 257, "eligible": 1124, "ties": 365}}
+{"average_precision": 0.43956936525087315, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "confusion": {"fn": 401, "fp": 959, "tn": 521, "tp": 741}, "domain": "security", "f1": 0.5214637579169599, "false_positives_per_true_positive": 1.2941970310391362, "model_id": "Contrastive-LM/CLM-v0.1-8B", "model_revision": "b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5", "precision": 0.43588235294117644, "protocol_id": "clm-native-choice-6144-v1", "recall": 0.648861646234676, "retrospective_split": true, "run_id": "clm-base", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 466, "eligible": 1124, "ties": 0}}
+{"average_precision": 0.4357752989578152, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "confusion": {"fn": 2, "fp": 1479, "tn": 1, "tp": 1140}, "domain": "security", "f1": 0.6062217495346982, "false_positives_per_true_positive": 1.2973684210526315, "model_id": "cisco-securebert2.0-code-vuln", "model_revision": "cisco-ai/SecureBERT2.0-code-vuln-detection", "precision": 0.4352806414662085, "protocol_id": "perch-security-state-4096-v1", "recall": 0.9982486865148862, "retrospective_split": true, "run_id": "securebert-base-normalized", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 564, "eligible": 1124, "ties": 3}}
+{"average_precision": 0.48938849226674325, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "confusion": {"fn": 388, "fp": 9, "tn": 471, "tp": 10}, "domain": "security", "f1": 0.047961630695443645, "false_positives_per_true_positive": 0.9, "model_id": "jev", "model_revision": "hosted-1.13.0", "precision": 0.5263157894736842, "protocol_id": "perch-broad-v1", "recall": 0.02512562814070352, "retrospective_split": true, "run_id": "jev-1.13.0", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 144, "eligible": 398, "ties": 92}}
+{"average_precision": 0.4744779913481638, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "confusion": {"fn": 380, "fp": 22, "tn": 458, "tp": 18}, "domain": "security", "f1": 0.0821917808219178, "false_positives_per_true_positive": 1.2222222222222223, "model_id": "alibiserikbay/JevK5", "model_revision": "3c673298eb7f2dc7cb98019262c55b1fac01a0bc", "precision": 0.45, "protocol_id": "perch-broad-v1", "recall": 0.04522613065326633, "retrospective_split": true, "run_id": "jevk5-base", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 151, "eligible": 398, "ties": 57}}
+{"average_precision": 0.4728876281488367, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "confusion": {"fn": 391, "fp": 9, "tn": 471, "tp": 7}, "domain": "security", "f1": 0.033816425120772944, "false_positives_per_true_positive": 1.2857142857142858, "model_id": "vagmi/jev-lite", "model_revision": "gemma-4-E4B-it+jev-lite-released", "precision": 0.4375, "protocol_id": "perch-broad-v1", "recall": 0.017587939698492462, "retrospective_split": true, "run_id": "gemma-jev-lite-base", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 117, "eligible": 398, "ties": 132}}
+{"average_precision": 0.471468796194999, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "confusion": {"fn": 128, "fp": 306, "tn": 174, "tp": 270}, "domain": "security", "f1": 0.5544147843942505, "false_positives_per_true_positive": 1.1333333333333333, "model_id": "Contrastive-LM/CLM-v0.1-8B", "model_revision": "b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5", "precision": 0.46875, "protocol_id": "clm-native-choice-6144-v1", "recall": 0.678391959798995, "retrospective_split": true, "run_id": "clm-base", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 210, "eligible": 398, "ties": 0}}
+{"average_precision": 0.430399069654715, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "confusion": {"fn": 1, "fp": 478, "tn": 2, "tp": 397}, "domain": "security", "f1": 0.6237234878240377, "false_positives_per_true_positive": 1.2040302267002518, "model_id": "cisco-securebert2.0-code-vuln", "model_revision": "cisco-ai/SecureBERT2.0-code-vuln-detection", "precision": 0.45371428571428574, "protocol_id": "perch-security-state-4096-v1", "recall": 0.9974874371859297, "retrospective_split": true, "run_id": "securebert-base-normalized", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 222, "eligible": 398, "ties": 1}}

split-summary.json

@@ -0,0 +1,41 @@
+{
+  "private_repo_fraction": 0.2,
+  "private_repository_count": 348,
+  "repository_count": 1740,
+  "seed": 42,
+  "split_rule": "repository-grouped random split across both domains",
+  "tracks": {
+    "bug:private": {
+      "methods": 996,
+      "negative": 572,
+      "pairs": 498,
+      "positive": 424,
+      "repositories": 263,
+      "sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73"
+    },
+    "bug:public": {
+      "methods": 4440,
+      "negative": 2579,
+      "pairs": 2220,
+      "positive": 1861,
+      "repositories": 1044,
+      "sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2"
+    },
+    "security:private": {
+      "methods": 878,
+      "negative": 480,
+      "pairs": 439,
+      "positive": 398,
+      "repositories": 90,
+      "sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051"
+    },
+    "security:public": {
+      "methods": 2622,
+      "negative": 1480,
+      "pairs": 1311,
+      "positive": 1142,
+      "repositories": 371,
+      "sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8"
+    }
+  }
+}