Spaces:
Sleeping
Sleeping
| { | |
| "name": "Bash Only + Gemini 3.1 Pro", | |
| "model_id": "google/gemini-3.1-pro-preview", | |
| "tag": "rust-specific", | |
| "eval_filename": "2026-04-22T13-50-48-00-00_mizan-rust-specific_AkY5L29M6rrViMNc2NQdTQ.eval", | |
| "samples": [ | |
| { | |
| "example_id": "vuln-0001/sample-00001-crate", | |
| "vuln_id": "vuln-0001", | |
| "granularity": "crate", | |
| "crate_name": "chttp", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-415\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0001/sample-00001-file", | |
| "vuln_id": "vuln-0001", | |
| "granularity": "file", | |
| "crate_name": "chttp", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\", \"CWE-415\", \"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0001/sample-00001-function", | |
| "vuln_id": "vuln-0001", | |
| "granularity": "function", | |
| "crate_name": "chttp", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-415\", \"CWE-416\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0001/sample-10001-crate", | |
| "vuln_id": "vuln-0001", | |
| "granularity": "crate", | |
| "crate_name": "chttp", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0001/sample-10001-file", | |
| "vuln_id": "vuln-0001", | |
| "granularity": "file", | |
| "crate_name": "chttp", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0001/sample-10001-function", | |
| "vuln_id": "vuln-0001", | |
| "granularity": "function", | |
| "crate_name": "chttp", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0002/sample-00002-crate", | |
| "vuln_id": "vuln-0002", | |
| "granularity": "crate", | |
| "crate_name": "arc-swap", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-825\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 1, | |
| "line_tp": 1, | |
| "line_fp": 3, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0002/sample-00002-file", | |
| "vuln_id": "vuln-0002", | |
| "granularity": "file", | |
| "crate_name": "arc-swap", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 3, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0002/sample-00002-function", | |
| "vuln_id": "vuln-0002", | |
| "granularity": "function", | |
| "crate_name": "arc-swap", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-825\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 2, | |
| "line_tp": 1, | |
| "line_fp": 4, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0002/sample-10002-crate", | |
| "vuln_id": "vuln-0002", | |
| "granularity": "crate", | |
| "crate_name": "arc-swap", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0002/sample-10002-file", | |
| "vuln_id": "vuln-0002", | |
| "granularity": "file", | |
| "crate_name": "arc-swap", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0002/sample-10002-function", | |
| "vuln_id": "vuln-0002", | |
| "granularity": "function", | |
| "crate_name": "arc-swap", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0003/sample-00003-crate", | |
| "vuln_id": "vuln-0003", | |
| "granularity": "crate", | |
| "crate_name": "safe-transmute", | |
| "year": 2018, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 6, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0003/sample-00003-file", | |
| "vuln_id": "vuln-0003", | |
| "granularity": "file", | |
| "crate_name": "safe-transmute", | |
| "year": 2018, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 9, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 9, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0003/sample-00003-function", | |
| "vuln_id": "vuln-0003", | |
| "granularity": "function", | |
| "crate_name": "safe-transmute", | |
| "year": 2018, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-131\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0003/sample-10003-crate", | |
| "vuln_id": "vuln-0003", | |
| "granularity": "crate", | |
| "crate_name": "safe-transmute", | |
| "year": 2018, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-415\", \"CWE-681\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 6, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 6, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0003/sample-10003-file", | |
| "vuln_id": "vuln-0003", | |
| "granularity": "file", | |
| "crate_name": "safe-transmute", | |
| "year": 2018, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-415\", \"CWE-681\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 9, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 9, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0003/sample-10003-function", | |
| "vuln_id": "vuln-0003", | |
| "granularity": "function", | |
| "crate_name": "safe-transmute", | |
| "year": 2018, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-415\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0004/sample-00004-crate", | |
| "vuln_id": "vuln-0004", | |
| "granularity": "crate", | |
| "crate_name": "ordnung", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-415\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0004/sample-00004-file", | |
| "vuln_id": "vuln-0004", | |
| "granularity": "file", | |
| "crate_name": "ordnung", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-415\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-415\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0004/sample-00004-function", | |
| "vuln_id": "vuln-0004", | |
| "granularity": "function", | |
| "crate_name": "ordnung", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-415\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-415\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0005/sample-00005-crate", | |
| "vuln_id": "vuln-0005", | |
| "granularity": "crate", | |
| "crate_name": "image", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-400\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0005/sample-00005-file", | |
| "vuln_id": "vuln-0005", | |
| "granularity": "file", | |
| "crate_name": "image", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-400\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0005/sample-00005-function", | |
| "vuln_id": "vuln-0005", | |
| "granularity": "function", | |
| "crate_name": "image", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-400\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0005/sample-10005-crate", | |
| "vuln_id": "vuln-0005", | |
| "granularity": "crate", | |
| "crate_name": "image", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0005/sample-10005-file", | |
| "vuln_id": "vuln-0005", | |
| "granularity": "file", | |
| "crate_name": "image", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0005/sample-10005-function", | |
| "vuln_id": "vuln-0005", | |
| "granularity": "function", | |
| "crate_name": "image", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0006/sample-00006-crate", | |
| "vuln_id": "vuln-0006", | |
| "granularity": "crate", | |
| "crate_name": "smallvec", | |
| "year": 2021, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 1, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0006/sample-00006-file", | |
| "vuln_id": "vuln-0006", | |
| "granularity": "file", | |
| "crate_name": "smallvec", | |
| "year": 2021, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 2, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0006/sample-00006-function", | |
| "vuln_id": "vuln-0006", | |
| "granularity": "function", | |
| "crate_name": "smallvec", | |
| "year": 2021, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0006/sample-10006-crate", | |
| "vuln_id": "vuln-0006", | |
| "granularity": "crate", | |
| "crate_name": "smallvec", | |
| "year": 2021, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0006/sample-10006-file", | |
| "vuln_id": "vuln-0006", | |
| "granularity": "file", | |
| "crate_name": "smallvec", | |
| "year": 2021, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0006/sample-10006-function", | |
| "vuln_id": "vuln-0006", | |
| "granularity": "function", | |
| "crate_name": "smallvec", | |
| "year": 2021, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0007/sample-00007-crate", | |
| "vuln_id": "vuln-0007", | |
| "granularity": "crate", | |
| "crate_name": "ordnung", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-131\", \"CWE-119\", \"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 5, | |
| "function_fn": 17, | |
| "line_tp": 0, | |
| "line_fp": 5, | |
| "line_fn": 18, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0007/sample-00007-file", | |
| "vuln_id": "vuln-0007", | |
| "granularity": "file", | |
| "crate_name": "ordnung", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-415\", \"CWE-416\", \"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 17, | |
| "line_tp": 1, | |
| "line_fp": 2, | |
| "line_fn": 17, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0007/sample-00007-function", | |
| "vuln_id": "vuln-0007", | |
| "granularity": "function", | |
| "crate_name": "ordnung", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0008/sample-00008-crate", | |
| "vuln_id": "vuln-0008", | |
| "granularity": "crate", | |
| "crate_name": "magnetic", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 17, | |
| "function_fn": 13, | |
| "line_tp": 13, | |
| "line_fp": 4, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0008/sample-00008-function", | |
| "vuln_id": "vuln-0008", | |
| "granularity": "function", | |
| "crate_name": "magnetic", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-778\", \"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 20, | |
| "function_fn": 13, | |
| "line_tp": 13, | |
| "line_fp": 7, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0008/sample-10008-crate", | |
| "vuln_id": "vuln-0008", | |
| "granularity": "crate", | |
| "crate_name": "magnetic", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0008/sample-10008-function", | |
| "vuln_id": "vuln-0008", | |
| "granularity": "function", | |
| "crate_name": "magnetic", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-662\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0009/sample-00009-crate", | |
| "vuln_id": "vuln-0009", | |
| "granularity": "crate", | |
| "crate_name": "alg_ds", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-824\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0009/sample-00009-file", | |
| "vuln_id": "vuln-0009", | |
| "granularity": "file", | |
| "crate_name": "alg_ds", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0009/sample-00009-function", | |
| "vuln_id": "vuln-0009", | |
| "granularity": "function", | |
| "crate_name": "alg_ds", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0010/sample-00010-crate", | |
| "vuln_id": "vuln-0010", | |
| "granularity": "crate", | |
| "crate_name": "im", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0010/sample-00010-file", | |
| "vuln_id": "vuln-0010", | |
| "granularity": "file", | |
| "crate_name": "im", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0010/sample-00010-function", | |
| "vuln_id": "vuln-0010", | |
| "granularity": "function", | |
| "crate_name": "im", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0010/sample-10010-crate", | |
| "vuln_id": "vuln-0010", | |
| "granularity": "crate", | |
| "crate_name": "im", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0010/sample-10010-file", | |
| "vuln_id": "vuln-0010", | |
| "granularity": "file", | |
| "crate_name": "im", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0010/sample-10010-function", | |
| "vuln_id": "vuln-0010", | |
| "granularity": "function", | |
| "crate_name": "im", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-778\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0011/sample-00011-crate", | |
| "vuln_id": "vuln-0011", | |
| "granularity": "crate", | |
| "crate_name": "array-init-cursor", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-672\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-415\", \"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0011/sample-00011-function", | |
| "vuln_id": "vuln-0011", | |
| "granularity": "function", | |
| "crate_name": "array-init-cursor", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-672\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-415\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0011/sample-10011-crate", | |
| "vuln_id": "vuln-0011", | |
| "granularity": "crate", | |
| "crate_name": "array-init-cursor", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-843\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0011/sample-10011-function", | |
| "vuln_id": "vuln-0011", | |
| "granularity": "function", | |
| "crate_name": "array-init-cursor", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0012/sample-00012-crate", | |
| "vuln_id": "vuln-0012", | |
| "granularity": "crate", | |
| "crate_name": "hashconsing", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-662\", \"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 2, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0012/sample-00012-file", | |
| "vuln_id": "vuln-0012", | |
| "granularity": "file", | |
| "crate_name": "hashconsing", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-662\", \"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 2, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0012/sample-00012-function", | |
| "vuln_id": "vuln-0012", | |
| "granularity": "function", | |
| "crate_name": "hashconsing", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-662\", \"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 2, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0012/sample-10012-crate", | |
| "vuln_id": "vuln-0012", | |
| "granularity": "crate", | |
| "crate_name": "hashconsing", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0012/sample-10012-file", | |
| "vuln_id": "vuln-0012", | |
| "granularity": "file", | |
| "crate_name": "hashconsing", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0012/sample-10012-function", | |
| "vuln_id": "vuln-0012", | |
| "granularity": "function", | |
| "crate_name": "hashconsing", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0013/sample-00013-crate", | |
| "vuln_id": "vuln-0013", | |
| "granularity": "crate", | |
| "crate_name": "bitvec", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-415\", \"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 2, | |
| "function_tp": 0, | |
| "function_fp": 5, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 9, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0013/sample-00013-file", | |
| "vuln_id": "vuln-0013", | |
| "granularity": "file", | |
| "crate_name": "bitvec", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-415\", \"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 4, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 4, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0013/sample-00013-function", | |
| "vuln_id": "vuln-0013", | |
| "granularity": "function", | |
| "crate_name": "bitvec", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-415\", \"CWE-416\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0013/sample-10013-crate", | |
| "vuln_id": "vuln-0013", | |
| "granularity": "crate", | |
| "crate_name": "bitvec", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-824\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0013/sample-10013-file", | |
| "vuln_id": "vuln-0013", | |
| "granularity": "file", | |
| "crate_name": "bitvec", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0013/sample-10013-function", | |
| "vuln_id": "vuln-0013", | |
| "granularity": "function", | |
| "crate_name": "bitvec", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0014/sample-00014-crate", | |
| "vuln_id": "vuln-0014", | |
| "granularity": "crate", | |
| "crate_name": "redox_uefi_std", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-122\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-476\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 1, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0014/sample-00014-file", | |
| "vuln_id": "vuln-0014", | |
| "granularity": "file", | |
| "crate_name": "redox_uefi_std", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-122\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-285\", \"CWE-476\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0014/sample-00014-function", | |
| "vuln_id": "vuln-0014", | |
| "granularity": "function", | |
| "crate_name": "redox_uefi_std", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-122\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 3, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0014/sample-10014-crate", | |
| "vuln_id": "vuln-0014", | |
| "granularity": "crate", | |
| "crate_name": "redox_uefi_std", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-284\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 6, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 6, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0014/sample-10014-file", | |
| "vuln_id": "vuln-0014", | |
| "granularity": "file", | |
| "crate_name": "redox_uefi_std", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-476\", \"CWE-362\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0014/sample-10014-function", | |
| "vuln_id": "vuln-0014", | |
| "granularity": "function", | |
| "crate_name": "redox_uefi_std", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-130\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0015/sample-00015-crate", | |
| "vuln_id": "vuln-0015", | |
| "granularity": "crate", | |
| "crate_name": "arr", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\", \"CWE-787\", \"CWE-362\", \"CWE-131\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 4, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 8, | |
| "function_fn": 2, | |
| "line_tp": 2, | |
| "line_fp": 6, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0015/sample-00015-function", | |
| "vuln_id": "vuln-0015", | |
| "granularity": "function", | |
| "crate_name": "arr", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0016/sample-00016-crate", | |
| "vuln_id": "vuln-0016", | |
| "granularity": "crate", | |
| "crate_name": "arr", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-120\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0016/sample-00016-function", | |
| "vuln_id": "vuln-0016", | |
| "granularity": "function", | |
| "crate_name": "arr", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-120\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\", \"CWE-787\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0017/sample-00017-crate", | |
| "vuln_id": "vuln-0017", | |
| "granularity": "crate", | |
| "crate_name": "arr", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\", \"CWE-787\", \"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 4, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 6, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 7, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0017/sample-00017-function", | |
| "vuln_id": "vuln-0017", | |
| "granularity": "function", | |
| "crate_name": "arr", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-824\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0018/sample-00018-crate", | |
| "vuln_id": "vuln-0018", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-682\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 4, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 4, | |
| "line_fn": 4, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0018/sample-00018-file", | |
| "vuln_id": "vuln-0018", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-787\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 12, | |
| "line_fn": 4, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0018/sample-00018-function", | |
| "vuln_id": "vuln-0018", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-787\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 3, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0018/sample-10018-crate", | |
| "vuln_id": "vuln-0018", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0018/sample-10018-file", | |
| "vuln_id": "vuln-0018", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0018/sample-10018-function", | |
| "vuln_id": "vuln-0018", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0019/sample-00019-crate", | |
| "vuln_id": "vuln-0019", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 6, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0019/sample-00019-file", | |
| "vuln_id": "vuln-0019", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-824\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 21, | |
| "line_fn": 6, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0019/sample-00019-function", | |
| "vuln_id": "vuln-0019", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 3, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0019/sample-10019-crate", | |
| "vuln_id": "vuln-0019", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0019/sample-10019-file", | |
| "vuln_id": "vuln-0019", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-415\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0019/sample-10019-function", | |
| "vuln_id": "vuln-0019", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0020/sample-00020-crate", | |
| "vuln_id": "vuln-0020", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0020/sample-00020-file", | |
| "vuln_id": "vuln-0020", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0020/sample-00020-function", | |
| "vuln_id": "vuln-0020", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0020/sample-10020-crate", | |
| "vuln_id": "vuln-0020", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0020/sample-10020-file", | |
| "vuln_id": "vuln-0020", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0020/sample-10020-function", | |
| "vuln_id": "vuln-0020", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-121\", \"CWE-682\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 6, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 6, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0021/sample-00021-crate", | |
| "vuln_id": "vuln-0021", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-401\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0021/sample-00021-file", | |
| "vuln_id": "vuln-0021", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-401\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-415\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0021/sample-00021-function", | |
| "vuln_id": "vuln-0021", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-401\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0021/sample-10021-crate", | |
| "vuln_id": "vuln-0021", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\", \"CWE-787\", \"CWE-119\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0021/sample-10021-file", | |
| "vuln_id": "vuln-0021", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-668\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0021/sample-10021-function", | |
| "vuln_id": "vuln-0021", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0022/sample-00022-crate", | |
| "vuln_id": "vuln-0022", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-401\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0022/sample-00022-file", | |
| "vuln_id": "vuln-0022", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-401\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0022/sample-00022-function", | |
| "vuln_id": "vuln-0022", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-401\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\", \"CWE-824\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0022/sample-10022-crate", | |
| "vuln_id": "vuln-0022", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0022/sample-10022-file", | |
| "vuln_id": "vuln-0022", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-668\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0022/sample-10022-function", | |
| "vuln_id": "vuln-0022", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0023/sample-00023-crate", | |
| "vuln_id": "vuln-0023", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0023/sample-00023-file", | |
| "vuln_id": "vuln-0023", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-787\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 8, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 8, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0023/sample-00023-function", | |
| "vuln_id": "vuln-0023", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-129\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-190\", \"CWE-787\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 8, | |
| "line_tp": 0, | |
| "line_fp": 7, | |
| "line_fn": 8, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0023/sample-10023-crate", | |
| "vuln_id": "vuln-0023", | |
| "granularity": "crate", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0023/sample-10023-file", | |
| "vuln_id": "vuln-0023", | |
| "granularity": "file", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0023/sample-10023-function", | |
| "vuln_id": "vuln-0023", | |
| "granularity": "function", | |
| "crate_name": "sized-chunks", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0024/sample-00024-crate", | |
| "vuln_id": "vuln-0024", | |
| "granularity": "crate", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\", \"CWE-908\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0024/sample-00024-function", | |
| "vuln_id": "vuln-0024", | |
| "granularity": "function", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0024/sample-10024-crate", | |
| "vuln_id": "vuln-0024", | |
| "granularity": "crate", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-190\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0024/sample-10024-function", | |
| "vuln_id": "vuln-0024", | |
| "granularity": "function", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0025/sample-00025-crate", | |
| "vuln_id": "vuln-0025", | |
| "granularity": "crate", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-193\", \"CWE-908\", \"CWE-401\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\", \"CWE-457\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 3, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0025/sample-00025-function", | |
| "vuln_id": "vuln-0025", | |
| "granularity": "function", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-193\", \"CWE-908\", \"CWE-401\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 3, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0025/sample-10025-crate", | |
| "vuln_id": "vuln-0025", | |
| "granularity": "crate", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-190\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 3, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 4, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0025/sample-10025-function", | |
| "vuln_id": "vuln-0025", | |
| "granularity": "function", | |
| "crate_name": "simple-slab", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-415\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0026/sample-00026-crate", | |
| "vuln_id": "vuln-0026", | |
| "granularity": "crate", | |
| "crate_name": "vmm-sys-util", | |
| "year": 2023, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0026/sample-00026-file", | |
| "vuln_id": "vuln-0026", | |
| "granularity": "file", | |
| "crate_name": "vmm-sys-util", | |
| "year": 2023, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-200\", \"CWE-908\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0026/sample-10026-crate", | |
| "vuln_id": "vuln-0026", | |
| "granularity": "crate", | |
| "crate_name": "vmm-sys-util", | |
| "year": 2023, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0026/sample-10026-file", | |
| "vuln_id": "vuln-0026", | |
| "granularity": "file", | |
| "crate_name": "vmm-sys-util", | |
| "year": 2023, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0027/sample-00027-crate", | |
| "vuln_id": "vuln-0027", | |
| "granularity": "crate", | |
| "crate_name": "claxon", | |
| "year": 2018, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 1, | |
| "line_fp": 1, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0027/sample-00027-file", | |
| "vuln_id": "vuln-0027", | |
| "granularity": "file", | |
| "crate_name": "claxon", | |
| "year": 2018, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-908\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 2, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0027/sample-10027-crate", | |
| "vuln_id": "vuln-0027", | |
| "granularity": "crate", | |
| "crate_name": "claxon", | |
| "year": 2018, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0027/sample-10027-file", | |
| "vuln_id": "vuln-0027", | |
| "granularity": "file", | |
| "crate_name": "claxon", | |
| "year": 2018, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-908\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0028/sample-00028-crate", | |
| "vuln_id": "vuln-0028", | |
| "granularity": "crate", | |
| "crate_name": "once_cell", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-20\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0028/sample-00028-function", | |
| "vuln_id": "vuln-0028", | |
| "granularity": "function", | |
| "crate_name": "once_cell", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-20\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-617\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0028/sample-10028-crate", | |
| "vuln_id": "vuln-0028", | |
| "granularity": "crate", | |
| "crate_name": "once_cell", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0028/sample-10028-function", | |
| "vuln_id": "vuln-0028", | |
| "granularity": "function", | |
| "crate_name": "once_cell", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0029/sample-00029-crate", | |
| "vuln_id": "vuln-0029", | |
| "granularity": "crate", | |
| "crate_name": "pared", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 6, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 8, | |
| "line_fn": 6, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0029/sample-00029-function", | |
| "vuln_id": "vuln-0029", | |
| "granularity": "function", | |
| "crate_name": "pared", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0029/sample-10029-crate", | |
| "vuln_id": "vuln-0029", | |
| "granularity": "crate", | |
| "crate_name": "pared", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-415\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 12, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 12, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0029/sample-10029-function", | |
| "vuln_id": "vuln-0029", | |
| "granularity": "function", | |
| "crate_name": "pared", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0030/sample-00030-crate", | |
| "vuln_id": "vuln-0030", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0030/sample-10030-crate", | |
| "vuln_id": "vuln-0030", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0031/sample-00031-crate", | |
| "vuln_id": "vuln-0031", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 4, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0031/sample-10031-crate", | |
| "vuln_id": "vuln-0031", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0032/sample-00032-crate", | |
| "vuln_id": "vuln-0032", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 3, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0032/sample-10032-crate", | |
| "vuln_id": "vuln-0032", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0033/sample-00033-crate", | |
| "vuln_id": "vuln-0033", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-134\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\", \"CWE-134\", \"CWE-362\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 14, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 15, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0033/sample-10033-crate", | |
| "vuln_id": "vuln-0033", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0034/sample-00034-crate", | |
| "vuln_id": "vuln-0034", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0034/sample-10034-crate", | |
| "vuln_id": "vuln-0034", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 12, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 9, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0035/sample-00035-crate", | |
| "vuln_id": "vuln-0035", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-362\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 4, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0035/sample-10035-crate", | |
| "vuln_id": "vuln-0035", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0036/sample-00036-crate", | |
| "vuln_id": "vuln-0036", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-351\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-362\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 4, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 4, | |
| "line_fn": 3, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0036/sample-10036-crate", | |
| "vuln_id": "vuln-0036", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0037/sample-00037-crate", | |
| "vuln_id": "vuln-0037", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-416\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-134\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0037/sample-10037-crate", | |
| "vuln_id": "vuln-0037", | |
| "granularity": "crate", | |
| "crate_name": "rusqlite", | |
| "year": 2020, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0038/sample-00038-crate", | |
| "vuln_id": "vuln-0038", | |
| "granularity": "crate", | |
| "crate_name": "scanner", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\", \"CWE-787\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 1, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0038/sample-00038-function", | |
| "vuln_id": "vuln-0038", | |
| "granularity": "function", | |
| "crate_name": "scanner", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 2, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0039/sample-00039-crate", | |
| "vuln_id": "vuln-0039", | |
| "granularity": "crate", | |
| "crate_name": "smallvec", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\", \"CWE-908\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 3, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 3, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0039/sample-00039-function", | |
| "vuln_id": "vuln-0039", | |
| "granularity": "function", | |
| "crate_name": "smallvec", | |
| "year": 2019, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-787\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-843\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0039/sample-10039-crate", | |
| "vuln_id": "vuln-0039", | |
| "granularity": "crate", | |
| "crate_name": "smallvec", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0039/sample-10039-function", | |
| "vuln_id": "vuln-0039", | |
| "granularity": "function", | |
| "crate_name": "smallvec", | |
| "year": 2019, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0040/sample-00040-crate", | |
| "vuln_id": "vuln-0040", | |
| "granularity": "crate", | |
| "crate_name": "mp3-metadata", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 1, | |
| "function_tp": 1, | |
| "function_fp": 1, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 22, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0040/sample-00040-function", | |
| "vuln_id": "vuln-0040", | |
| "granularity": "function", | |
| "crate_name": "mp3-metadata", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 1, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 1, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0040/sample-10040-crate", | |
| "vuln_id": "vuln-0040", | |
| "granularity": "crate", | |
| "crate_name": "mp3-metadata", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0040/sample-10040-function", | |
| "vuln_id": "vuln-0040", | |
| "granularity": "function", | |
| "crate_name": "mp3-metadata", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": false, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0041/sample-00041-crate", | |
| "vuln_id": "vuln-0041", | |
| "granularity": "crate", | |
| "crate_name": "xmas-elf", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-416\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 1, | |
| "function_tp": 0, | |
| "function_fp": 5, | |
| "function_fn": 2, | |
| "line_tp": 0, | |
| "line_fp": 5, | |
| "line_fn": 2, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0041/sample-00041-function", | |
| "vuln_id": "vuln-0041", | |
| "granularity": "function", | |
| "crate_name": "xmas-elf", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-125\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 2, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 2, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0041/sample-10041-crate", | |
| "vuln_id": "vuln-0041", | |
| "granularity": "crate", | |
| "crate_name": "xmas-elf", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\", \"CWE-758\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0041/sample-10041-function", | |
| "vuln_id": "vuln-0041", | |
| "granularity": "function", | |
| "crate_name": "xmas-elf", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\"]", | |
| "binary_accuracy": 0.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 2, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 2, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0042/sample-00042-crate", | |
| "vuln_id": "vuln-0042", | |
| "granularity": "crate", | |
| "crate_name": "fast-float", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 7, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 8, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0042/sample-00042-file", | |
| "vuln_id": "vuln-0042", | |
| "granularity": "file", | |
| "crate_name": "fast-float", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-119\", \"CWE-125\", \"CWE-787\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 2, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 7, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 8, | |
| "line_fn": 1, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0042/sample-00042-function", | |
| "vuln_id": "vuln-0042", | |
| "granularity": "function", | |
| "crate_name": "fast-float", | |
| "year": 2025, | |
| "is_vulnerable_gt": true, | |
| "cwe_types_gt": "[\"CWE-119\"]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": true, | |
| "cwe_types_pred": "[\"CWE-125\", \"CWE-119\"]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 1, | |
| "cwe_fp": 1, | |
| "cwe_fn": 0, | |
| "function_tp": 1, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 1, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 1.0, | |
| "success_at_1_line": 1.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0042/sample-10042-crate", | |
| "vuln_id": "vuln-0042", | |
| "granularity": "crate", | |
| "crate_name": "fast-float", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0042/sample-10042-file", | |
| "vuln_id": "vuln-0042", | |
| "granularity": "file", | |
| "crate_name": "fast-float", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| }, | |
| { | |
| "example_id": "vuln-0042/sample-10042-function", | |
| "vuln_id": "vuln-0042", | |
| "granularity": "function", | |
| "crate_name": "fast-float", | |
| "year": 2025, | |
| "is_vulnerable_gt": false, | |
| "cwe_types_gt": "[]", | |
| "is_valid_json": true, | |
| "is_vulnerable_pred": false, | |
| "cwe_types_pred": "[]", | |
| "binary_accuracy": 1.0, | |
| "cwe_tp": 0, | |
| "cwe_fp": 0, | |
| "cwe_fn": 0, | |
| "function_tp": 0, | |
| "function_fp": 0, | |
| "function_fn": 0, | |
| "line_tp": 0, | |
| "line_fp": 0, | |
| "line_fn": 0, | |
| "success_at_1_function": 0.0, | |
| "success_at_1_line": 0.0, | |
| "model": "Bash Only + Gemini 3.1 Pro" | |
| } | |
| ] | |
| } |