{
  "version": "devtools-v1",
  "seed": "devtools-v1",
  "partitions": [
    "train",
    "development",
    "test"
  ],
  "locked": [
    "test"
  ],
  "files": {
    "train.jsonl": {
      "sha256": "587682c145cbbc0c8c990e9d649da7b972853a9eaaef71e9b7b0fa18163b962c",
      "records": 5320,
      "questions": 6973,
      "bytes": 8924680,
      "in_git": true,
      "by_source": {
        "aegis": 1500,
        "codereviewer": 1500,
        "commitpackft": 1500,
        "flakeflagger": 820
      }
    },
    "development.jsonl": {
      "sha256": "3df3aa83890c43db65ac73b5bb6e4fbb1466362fc1a25dbd383d4d2edd0a1fe6",
      "records": 900,
      "questions": 1074,
      "bytes": 1731002,
      "in_git": true,
      "by_source": {
        "aegis": 150,
        "codereviewer": 150,
        "commitpackft": 150,
        "flakeflagger": 150,
        "prompt_injection": 150,
        "when2call": 150
      }
    },
    "test.jsonl": {
      "sha256": "2fb7c2a239fef923bcb62c0eee9d6884a608ce30e5e74a4e6b9082787679767f",
      "records": 900,
      "questions": 1073,
      "bytes": 1686530,
      "in_git": true,
      "by_source": {
        "aegis": 150,
        "codereviewer": 150,
        "commitpackft": 150,
        "flakeflagger": 150,
        "prompt_injection": 150,
        "when2call": 150
      }
    }
  },
  "trainable_sources": [
    "codereviewer",
    "commitpackft",
    "aegis",
    "flakeflagger"
  ],
  "eval_only_sources": [
    "when2call",
    "prompt_injection"
  ],
  "holdout_sources": [],
  "sources": {
    "codereviewer": {
      "trainable": true,
      "licence": "CC-BY-4.0",
      "licence_url": "https://zenodo.org/records/6900648",
      "attribution": "Li et al., Automating Code Review Activities by Large-Scale Pre-Training (ESEC/FSE 2022), Microsoft CodeReviewer",
      "label_provenance": "human: y = 1 when a human reviewer left a comment on the hunk in the original GitHub pull request, 0 otherwise (dataset label, unchanged)",
      "licence_check": "diffs come from GitHub projects: kept only projects whose current GitHub licence (GitHub API spdx_id, scripts/devtools_v1_licences.json) is permissive; GPL/AGPL/LGPL/MPL/OSL/MS-PL, NOASSERTION and no-licence projects dropped. The licence is the repository's at lookup time, not at collection time.",
      "pins": {
        "zenodo": {
          "record": 6900648,
          "doi": "10.5281/zenodo.6900648",
          "file": "Diff_Quality_Estimation.zip",
          "md5": "aad78e57d7d591172922da96e38e47dd",
          "sha256": "86d054de47741cb358c8a17ab55b6191356fc63e44010c864dd790481f41fff5",
          "members": [
            "cls-valid.jsonl",
            "cls-test.jsonl"
          ]
        },
        "licences": "scripts/devtools_v1_licences.json#codereviewer"
      }
    },
    "commitpackft": {
      "trainable": true,
      "licence": "MIT (dataset); per-row repository licence",
      "licence_url": "https://huggingface.co/datasets/bigcode/commitpackft",
      "attribution": "Muennighoff et al., OctoPack: Instruction Tuning Code Large Language Models (2023), BigCode",
      "label_provenance": "heuristic: the commit type is the class of the first verb of the author's commit subject (commit_type); message match is by construction (the commit's own subject vs another commit's subject)",
      "licence_check": "rows kept only when the per-row `license` (from the GitHub repository) is one of mit, apache-2.0, bsd-2-clause, bsd-3-clause, isc, cc0-1.0, unlicense",
      "pins": {
        "hf": "bigcode/commitpackft@fc56fe33c030c6daa414c2b112c932b8eed085e6",
        "files": {
          "data/python/data.jsonl": "d167da37e1058371c48e057cd8815d03700c867dd8bcf58e61420d4dcd288d73",
          "data/javascript/data.jsonl": "968688a1163c2dbbced0d0f9cf1dd52bc0b626c2ad9b02ff23f5700b2e277f25",
          "data/typescript/data.jsonl": "43d32596844c62aa23b0e5a8811648f6f20fa07d9ab80499d66507136b844a76",
          "data/java/data.jsonl": "c51a331320f4220d6053e3389485916c738fcaccfd583298b7738af834fa19b9",
          "data/go/data.jsonl": "05a65e973fcc6894a234fb382987cc25e26f053d60b032ea83540a5b36a528b8",
          "data/rust/data.jsonl": "5611f3701026acf3c43ba64da50f4f65a078e84fba56e2bf313e8f2996b0cc68",
          "data/c/data.jsonl": "a67d9aeac62e693fb782030ad7c4aa9e0d9f38a02bd2e06a438604a92628ca71",
          "data/c++/data.jsonl": "ee59eb0dca07dadb74bde829134e0d3c42176e198773f125a14554a5547df4f2",
          "data/c#/data.jsonl": "a58a0a03a9965fae11e20c53cf5661ecca6ae5b3ca241f2989db4c2c1be5c04a",
          "data/ruby/data.jsonl": "6b1970e1c286f98445e519b92360eac5618f29f422c56ac949623d42dccd04b1",
          "data/php/data.jsonl": "fda88bbb6483a504dcbaec36abd363eb6ae99531431fecc924d95ecdad91c691",
          "data/shell/data.jsonl": "08eef5271d29368db2af8b0ca127c8d249d056ae7f6bce9c0692ade5dcad66ca"
        }
      }
    },
    "aegis": {
      "trainable": true,
      "licence": "CC-BY-4.0",
      "licence_url": "https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0",
      "attribution": "Ghosh et al., AEGIS2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails (NAACL 2025), NVIDIA",
      "label_provenance": "human: prompt_label (prompt_label_source is `human` on every row); category from the human violated_categories of prompt-only rows. Responses (Mistral-7B text, partly LLM-jury labels) are not used",
      "licence_check": "card licence CC-BY-4.0, not gated. Prompts come from Anthropic HH-RLHF (MIT), DAN / jailbreak_llms (MIT) and AART (PaLM-generated): AART matches dropped. Note the card's out-of-scope paragraph: 'The data are intended for research purposes, especially research that can make models less harmful' next to the CC-BY-4.0 licence and the stated direct use 'to build LLM content safety moderation guardrails'",
      "pins": {
        "hf": "nvidia/Aegis-AI-Content-Safety-Dataset-2.0@d86bb8bedff51d25ac834ab7838f1cc61acb7a2c",
        "files": {
          "train.json": "154fba82c71d9fa73abd2ca5588a198e693ddc816c83444df180a22f613e02f6",
          "validation.json": "a97200e226ad4f6ba6a639982f817f675909bc74513859df3e8fc9a92951dfcd",
          "test.json": "b0a6d602260524866053cb34105194f074f2c2906e3691b68d43b9e6e9318f35"
        },
        "aart_filter": {
          "hf": "walledai/AART@dd98454c52d34e860f5927625a3989edd5617d78",
          "files": {
            "data/train-00000-of-00001.parquet": "d22e3a35637a129bbffa6055a9ae83df1327f373ea22ef6f3cfa1c2d17eba0f1"
          }
        }
      }
    },
    "when2call": {
      "trainable": false,
      "licence": "CC-BY-4.0",
      "licence_url": "https://huggingface.co/datasets/nvidia/When2Call",
      "attribution": "Ross et al., When2Call: When (not) to Call Tools (NAACL 2025), NVIDIA",
      "label_provenance": "by construction (NVIDIA's pipeline over BFCL v2 Live): tool_call = the target tool is available, cannot_answer = the target tool was removed, request_for_info = a required parameter was removed from the question (question rewritten by Mixtral-8x22B, an open model). `direct` is never correct in the test MCQ",
      "licence_check": "eval-only: only the test MCQ is used; it is derived from BFCL v2 Live (Apache-2.0), a community benchmark that must stay uncontaminated. The train splits derive from Salesforce/xlam-function-calling-60k, whose card carries a research-only statement, and are not used",
      "pins": {
        "hf": "nvidia/When2Call@0582f7749df63a96fdc3070932e83e72396ace53",
        "files": {
          "test/when2call_test_mcq.jsonl": "8c3694e583eeeb8dbc297e6cd90da70efc68efa4b6adb7227523e828c6b7b14c"
        }
      }
    },
    "flakeflagger": {
      "trainable": true,
      "licence": "CC-BY-4.0 (Zenodo data); BSD-3-Clause (GitHub repo with the test bodies); test code under each project's licence",
      "licence_url": "https://zenodo.org/records/4450723",
      "attribution": "Alshammari et al., FlakeFlagger: Predicting Flakiness Without Rerunning Tests (ICSE 2021)",
      "label_provenance": "heuristic (execution): IsFlaky from test_results.csv, i.e. the test both passed and failed across 10,000 reruns of its suite at the pinned commit",
      "licence_check": "each project's LICENSE at the commit in Project_Info.csv was read (scripts/devtools_v1_licences.json); wildfly (LGPL-2.1) and logback (EPL-1.0 / LGPL-2.1) dropped; togglz dropped (not in Project_Info.csv, no pinned commit)",
      "pins": {
        "zenodo": {
          "record": 4450723,
          "doi": "10.5281/zenodo.4450723",
          "files": {
            "test_results.csv": "86210ed8ac0171a3d64cf5ab83d503cc97e82845e0299b9f55599a8414218f13",
            "Project_Info.csv": "f0064b25ed64995842465b67c6fec86b5ddf72d30e1a7dc5b64abd8b0208e9fd"
          }
        },
        "github": "AlshammariA/FlakeFlagger@2fcbafc4713abbcc452bee07aa9681c7c8ccb707",
        "tarball_sha256": "67da177462a3026655a9ed8ef4e7c87a86d538c599ed484080db3b355f6d76b1",
        "licences": "scripts/devtools_v1_licences.json#flakeflagger"
      }
    },
    "prompt_injection": {
      "trainable": false,
      "licence": "Apache-2.0 (deepset/prompt-injections); MIT (Lakera/gandalf_ignore_instructions)",
      "licence_url": "https://huggingface.co/datasets/deepset/prompt-injections ; https://huggingface.co/datasets/Lakera/gandalf_ignore_instructions",
      "attribution": "deepset; Lakera (Gandalf)",
      "label_provenance": "deepset: dataset label (labeller not stated on the card). Gandalf: every row is an attack prompt submitted to the Gandalf password game, selected by OpenAI text-embedding-ada-002 similarity >= 0.825 to 'Ignore all previous instructions' (positives only). Negatives are deepset benign prompts only",
      "licence_check": "eval-only: small (662 + 1,000 rows), unstated labeller, closed-embedding selection of the attacks",
      "pins": {
        "deepset": {
          "hf": "deepset/prompt-injections@4f61ecb038e9c3fb77e21034b22511b523772cdd",
          "files": {
            "data/train-00000-of-00001-9564e8b05b4757ab.parquet": "2e10bc7ab30f542c97e4e83e2a5683000b5057d25ec10908784c631d44124c04",
            "data/test-00000-of-00001-701d16158af87368.parquet": "39ac797cabc157eeed58435a08593b2952bb6cb16fc394a2d383f447cc7b246e"
          }
        },
        "gandalf": {
          "hf": "Lakera/gandalf_ignore_instructions@04737b65e90a6794ec227012e4a255a7def6344b",
          "files": {
            "data/train-00000-of-00001-ded53be747ff55cd.parquet": "5b6acf3e5a5998d21f8e1222bb45bbdec25a14408747b1cd63bebef4a75fa439",
            "data/validation-00000-of-00001-94481a2a09ff2fff.parquet": "f51ab3e3407a368845b0f57932cc745c09280429416d7507bd178a16326a79f6",
            "data/test-00000-of-00001-bc92128b9288a6d1.parquet": "56b646d133335ebc535266bd55dbe1b5bee7caa4b95bf49d040684b9b5dd9972"
          }
        }
      }
    }
  },
  "counts": {
    "test": {
      "aegis": 150,
      "codereviewer": 150,
      "commitpackft": 150,
      "flakeflagger": 150,
      "prompt_injection": 150,
      "when2call": 150
    },
    "development": {
      "aegis": 150,
      "codereviewer": 150,
      "commitpackft": 150,
      "flakeflagger": 150,
      "prompt_injection": 150,
      "when2call": 150
    },
    "train": {
      "aegis": 1500,
      "codereviewer": 1500,
      "commitpackft": 1500,
      "flakeflagger": 820
    }
  },
  "groups": {
    "test": {
      "aegis": 150,
      "codereviewer": 16,
      "commitpackft": 141,
      "flakeflagger": 4,
      "prompt_injection": 150,
      "when2call": 135
    },
    "development": {
      "aegis": 150,
      "codereviewer": 10,
      "commitpackft": 137,
      "flakeflagger": 3,
      "prompt_injection": 150,
      "when2call": 132
    },
    "train": {
      "aegis": 1500,
      "codereviewer": 108,
      "commitpackft": 1374,
      "flakeflagger": 13
    }
  },
  "label_balance": {
    "test": {
      "aegis": {
        "aegis_category": {
          "criminal_planning": 5,
          "hate": 7,
          "pii": 4,
          "self_harm": 5,
          "sexual": 4
        },
        "aegis_unsafe": {
          "false": 75,
          "true": 75
        }
      },
      "codereviewer": {
        "codereviewer_needs_comment": {
          "false": 75,
          "true": 75
        }
      },
      "commitpackft": {
        "commitpackft_message": {
          "false": 74,
          "true": 74
        },
        "commitpackft_type": {
          "change": 22,
          "docs": 22,
          "feature": 22,
          "fix": 21,
          "remove": 21,
          "revert": 21,
          "test": 21
        }
      },
      "flakeflagger": {
        "flakeflagger_flaky": {
          "false": 75,
          "true": 75
        }
      },
      "prompt_injection": {
        "prompt_injection": {
          "false": 75,
          "true": 75
        }
      },
      "when2call": {
        "when2call_action": {
          "cannot_answer": 50,
          "request_for_info": 50,
          "tool_call": 50
        }
      }
    },
    "development": {
      "aegis": {
        "aegis_category": {
          "criminal_planning": 10,
          "harassment": 2,
          "hate": 2,
          "pii": 5,
          "profanity": 2,
          "self_harm": 2,
          "violence": 2
        },
        "aegis_unsafe": {
          "false": 75,
          "true": 75
        }
      },
      "codereviewer": {
        "codereviewer_needs_comment": {
          "false": 75,
          "true": 75
        }
      },
      "commitpackft": {
        "commitpackft_message": {
          "false": 75,
          "true": 74
        },
        "commitpackft_type": {
          "change": 22,
          "docs": 22,
          "feature": 22,
          "fix": 21,
          "remove": 21,
          "revert": 21,
          "test": 21
        }
      },
      "flakeflagger": {
        "flakeflagger_flaky": {
          "false": 75,
          "true": 75
        }
      },
      "prompt_injection": {
        "prompt_injection": {
          "false": 75,
          "true": 75
        }
      },
      "when2call": {
        "when2call_action": {
          "cannot_answer": 50,
          "request_for_info": 50,
          "tool_call": 50
        }
      }
    },
    "train": {
      "aegis": {
        "aegis_category": {
          "criminal_planning": 35,
          "harassment": 15,
          "hate": 14,
          "pii": 24,
          "profanity": 14,
          "self_harm": 15,
          "sexual": 18,
          "sexual_minor": 1,
          "substances": 6,
          "violence": 5,
          "weapons": 6
        },
        "aegis_unsafe": {
          "false": 750,
          "true": 750
        }
      },
      "codereviewer": {
        "codereviewer_needs_comment": {
          "false": 750,
          "true": 750
        }
      },
      "commitpackft": {
        "commitpackft_message": {
          "false": 750,
          "true": 750
        },
        "commitpackft_type": {
          "change": 215,
          "docs": 215,
          "feature": 214,
          "fix": 214,
          "remove": 214,
          "revert": 214,
          "test": 214
        }
      },
      "flakeflagger": {
        "flakeflagger_flaky": {
          "false": 410,
          "true": 410
        }
      }
    }
  },
  "state_tokens": {
    "test": {
      "aegis": {
        "min": 2,
        "median": 17,
        "p95": 362,
        "max": 989,
        "over_384": 8
      },
      "codereviewer": {
        "min": 84,
        "median": 232,
        "p95": 497,
        "max": 629,
        "over_384": 22
      },
      "commitpackft": {
        "min": 52,
        "median": 150,
        "p95": 561,
        "max": 835,
        "over_384": 20
      },
      "flakeflagger": {
        "min": 39,
        "median": 103,
        "p95": 398,
        "max": 1122,
        "over_384": 10
      },
      "prompt_injection": {
        "min": 4,
        "median": 14,
        "p95": 53,
        "max": 128,
        "over_384": 0
      },
      "when2call": {
        "min": 12,
        "median": 623,
        "p95": 1701,
        "max": 2521,
        "over_384": 102
      }
    },
    "development": {
      "aegis": {
        "min": 4,
        "median": 16,
        "p95": 402,
        "max": 2181,
        "over_384": 9
      },
      "codereviewer": {
        "min": 77,
        "median": 248,
        "p95": 507,
        "max": 674,
        "over_384": 19
      },
      "commitpackft": {
        "min": 48,
        "median": 157,
        "p95": 586,
        "max": 939,
        "over_384": 14
      },
      "flakeflagger": {
        "min": 32,
        "median": 213,
        "p95": 730,
        "max": 2313,
        "over_384": 30
      },
      "prompt_injection": {
        "min": 3,
        "median": 13,
        "p95": 62,
        "max": 162,
        "over_384": 0
      },
      "when2call": {
        "min": 13,
        "median": 467,
        "p95": 1586,
        "max": 2266,
        "over_384": 85
      }
    },
    "train": {
      "aegis": {
        "min": 2,
        "median": 18,
        "p95": 451,
        "max": 3285,
        "over_384": 93
      },
      "codereviewer": {
        "min": 31,
        "median": 234,
        "p95": 503,
        "max": 986,
        "over_384": 200
      },
      "commitpackft": {
        "min": 36,
        "median": 155,
        "p95": 542,
        "max": 947,
        "over_384": 190
      },
      "flakeflagger": {
        "min": 25,
        "median": 147,
        "p95": 667,
        "max": 2888,
        "over_384": 125
      }
    }
  },
  "tokenizer": {
    "model": "Qwen/Qwen3.5-4B-Base",
    "revision": "1001bb4d826a52d1f399e183466143f4da7b741b"
  },
  "base_revisions": {
    "Qwen/Qwen3.5-4B-Base": "1001bb4d826a52d1f399e183466143f4da7b741b"
  },
  "context": {
    "max_state": 7552,
    "max_branch": 8192,
    "max_packed": 9216,
    "truncate": false,
    "default_training_context": {
      "max_state": 384,
      "max_branch": 1024,
      "max_packed": 2048,
      "truncate": false
    },
    "note": "every record fits kev.model.training_context(MAX_TRAIN_STATE) under the tokenizer above; records whose state exceeds the default 384 tokens (state_tokens.*.over_<MAX_STATE>) need kev.train --max_state"
  },
  "selection": "development/test 150 records per source, train up to 1500 per trainable source; groups (repository, project, BFCL item, prompt) dealt to splits in a seeded hash order and never span splits; normalised-text state dedupe across all sources and splits (test, then development, then train); noul labels exactly balanced in pairs (within project for codereviewer / flakeflagger); choices drawn round-robin over labels; aegis keeps its native train/validation/test split; balance is a sampling choice, not the natural rate (flaky tests are 3.6% of FlakeFlagger)",
  "label_protocol": "no LLM labels: human (codereviewer, aegis), heuristic (commitpackft verb class, flakeflagger reruns), by construction (commitpackft message match, when2call), source label (prompt_injection)",
  "build_report": {
    "codereviewer": {
      "rows": 62504,
      "licence_dropped": 18917,
      "patch_size_dropped": 4,
      "candidates": 43583,
      "duplicate_state_within_source": 8466
    },
    "commitpackft": {
      "rows": 291782,
      "verb_unmapped": 48585,
      "licence_dropped": 23855,
      "diff_size_dropped": 279,
      "candidates": 219063,
      "by_type": {
        "change": 57242,
        "docs": 2657,
        "feature": 82243,
        "fix": 39254,
        "remove": 17640,
        "revert": 3757,
        "test": 16270
      },
      "duplicate_state_within_source": 4854
    },
    "aegis": {
      "rows": 28216,
      "redacted_dropped": 992,
      "aart_dropped": 3224,
      "label_conflict_dropped": 212,
      "candidates": 22381,
      "with_category": 2021,
      "duplicate_state_within_source": 0
    },
    "when2call": {
      "rows": 3652,
      "candidates": 3652,
      "by_label": {
        "cannot_answer": 1295,
        "request_for_info": 1062,
        "tool_call": 1295
      },
      "duplicate_state_within_source": 23
    },
    "flakeflagger": {
      "files": 21802,
      "label_mismatch_dropped": 1,
      "licence_dropped": 1884,
      "candidates": 19917,
      "flaky": 762,
      "tarball_sha256": "67da177462a3026655a9ed8ef4e7c87a86d538c599ed484080db3b355f6d76b1",
      "by_project": {
        "Alluxio-alluxio": [
          116,
          71
        ],
        "activiti-activiti": [
          32,
          2012
        ],
        "apache-ambari": [
          52,
          272
        ],
        "apache-commons-exec": [
          0,
          54
        ],
        "apache-hbase": [
          145,
          286
        ],
        "apache-httpcore": [
          22,
          690
        ],
        "apache-incubator-dubbo": [
          19,
          2157
        ],
        "doanduyhai-Achilles": [
          4,
          1313
        ],
        "elasticjob-elastic-job-lite": [
          3,
          555
        ],
        "hector-client-hector": [
          33,
          109
        ],
        "jknack-handlebars": [
          1,
          420
        ],
        "joel-costigliola-assertj-core": [
          1,
          6260
        ],
        "kevinsawicki-http-request": [
          18,
          145
        ],
        "ninjaframework-ninja": [
          1,
          306
        ],
        "orbit-orbit": [
          7,
          79
        ],
        "spring-projects-spring-boot": [
          160,
          1955
        ],
        "square-okhttp": [
          100,
          710
        ],
        "tootallnate-java-websocket": [
          23,
          122
        ],
        "undertow-io-undertow": [
          7,
          176
        ],
        "wro4j-wro4j": [
          16,
          1120
        ],
        "zxing-zxing": [
          2,
          343
        ]
      },
      "duplicate_state_within_source": 3108
    },
    "prompt_injection": {
      "duplicate_dropped": 1,
      "rows": 1662,
      "candidates": 1661,
      "by_origin_label": {
        "deepset/False": 399,
        "deepset/True": 263,
        "gandalf/True": 999
      },
      "duplicate_state_within_source": 0
    },
    "rejected": {
      "aegis/line_separator_rejected": 1
    },
    "pools": {
      "codereviewer": {
        "test": 3563,
        "development": 3539,
        "train": 28015
      },
      "commitpackft": {
        "test": 21421,
        "development": 21424,
        "train": 171364
      },
      "aegis": {
        "test": 1714,
        "development": 1028,
        "train": 19639
      },
      "when2call": {
        "test": 1814,
        "development": 1815,
        "train": 0
      },
      "flakeflagger": {
        "test": 1954,
        "development": 620,
        "train": 14235
      },
      "prompt_injection": {
        "test": 831,
        "development": 830,
        "train": 0
      }
    },
    "groups": {
      "codereviewer": {
        "test": 17,
        "development": 10,
        "train": 113
      },
      "commitpackft": {
        "test": 8063,
        "development": 7988,
        "train": 64340
      },
      "aegis": {
        "test": 1714,
        "development": 1028,
        "train": 19639
      },
      "when2call": {
        "test": 641,
        "development": 654,
        "train": 0
      },
      "flakeflagger": {
        "test": 4,
        "development": 3,
        "train": 14
      },
      "prompt_injection": {
        "test": 831,
        "development": 830,
        "train": 0
      }
    }
  },
  "licences_file": {
    "path": "scripts/devtools_v1_licences.json",
    "sha256": "75b45f29fe572f8cac0bcc97901b0a516797b9c4e32a281d867130b103abe3d9"
  },
  "large_partitions": "partitions with in_git false are gitignored; upload them to the Hub mirror (kev.suite.SUITES_DATASET) and bump SUITES_REVISION before load_split can fetch them elsewhere",
  "code_sha256": "ac336c1be8bb903f26760255743f30fe203cc535e48480029fe0680404c90b0d"
}