Datasets:
agent_ref dict | backend string | cwes list | family string | formal_target_id string | grading_mode string | grading_note string | instance dict | language string | lean_requirements dict | max_public_checks int64 | profile_id string | required_checks list | responses_create_params dict | row_schema_version int64 | slots list | source_provider string | source_target_id string | split_group string | subject string | task_id string | task_intent string | task_kind string | verifier_feedback string | vulnerability_class string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-122",
"CWE-787"
] | formal_task | OSS-FUZZ-394126186-poc-formal | legacy_lean_gate | released before per-profile validation commands; graded by the Lean gate alone (pinned obligation, kernel replay, axiom audit) | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-394126186-poc-formal.tar.zst",
"archive_sha256": "90d9db87f149d852cfd00ff500a44d0c698b18aedd530f43aa643c02b6c0d9aa",
"archive_size_bytes": 1236499,
"checker_image_id": null,
"formal_instance_digest": "4d6343e3056bdcb132564b9a49433c1ba63e02e883478fe82e58ea952de3e325... | C | {
"declarations": [],
"files": [
{
"allowed_imports": [
"Dependency",
"SourceModel",
"Spec",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogramming_allowed": false,
"module": "VulnerableDisproof",
"path": "submission/VulnerableDisproof.lean... | 20 | lean-translation-patch_validation-disproof-gen | [
"source_binding",
"formal_build",
"vulnerable_disproof",
"spec_fidelity"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "vulnerable_disproof",
"kind": "checked_witness",
"path": "submission/VulnerableDisproof.lean"
}
] | oss_fuzz | OSS-FUZZ-394126186-poc | tarantool/luajit | vulnerable | OSS-FUZZ-394126186-poc-formal::lean-translation-patch_validation-disproof-gen | patch_validation | disproof_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-122",
"CWE-787"
] | formal_task | OSS-FUZZ-394126186-poc-formal | not_gradable | the released backend has no per-profile validation commands; the released provenance lacks the patch-correspondence pointer; rebuild the instance with the current pipeline | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-394126186-poc-formal.tar.zst",
"archive_sha256": "90d9db87f149d852cfd00ff500a44d0c698b18aedd530f43aa643c02b6c0d9aa",
"archive_size_bytes": 1236499,
"checker_image_id": null,
"formal_instance_digest": "4d6343e3056bdcb132564b9a49433c1ba63e02e883478fe82e58ea952de3e325... | C | {
"declarations": [
{
"constant": "SecurityTarget.translation",
"module": "SourceModel"
}
],
"files": [
{
"allowed_imports": [
"Dependency",
"Lean",
"SourceModel",
"Spec",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogram... | 20 | lean-translation-patch_validation-e2e | [
"patch_allowlist",
"source_build",
"runtime_differential",
"formal_build",
"fixed_proof",
"vulnerable_disproof",
"spec_fidelity",
"assumption_audit"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "source_patch",
"kind": "source_patch",
"path": "submission/patch.diff"
},
{
"id": "specification",
"kind": "formal_specification",
"path": "submission/Spec.lean"
},
{
"id": "vulnerable_disproof",
"kind": "checked_witness",
"path": "submission/VulnerableDisproof.l... | oss_fuzz | OSS-FUZZ-394126186-poc | tarantool/luajit | pair | OSS-FUZZ-394126186-poc-formal::lean-translation-patch_validation-e2e | patch_validation | e2e | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-122",
"CWE-787"
] | formal_task | OSS-FUZZ-394126186-poc-formal | legacy_lean_gate | released before per-profile validation commands; graded by the Lean gate alone (pinned obligation, kernel replay, axiom audit) | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-394126186-poc-formal.tar.zst",
"archive_sha256": "90d9db87f149d852cfd00ff500a44d0c698b18aedd530f43aa643c02b6c0d9aa",
"archive_size_bytes": 1236499,
"checker_image_id": null,
"formal_instance_digest": "4d6343e3056bdcb132564b9a49433c1ba63e02e883478fe82e58ea952de3e325... | C | {
"declarations": [],
"files": [
{
"allowed_imports": [
"Dependency",
"Lean",
"SourceModel",
"Spec",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogramming_allowed": true,
"module": "FixedProof",
"path": "submission/FixedProof.lean"... | 20 | lean-translation-patch_validation-proof-gen | [
"source_binding",
"formal_build",
"fixed_proof",
"assumption_audit"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "fixed_proof",
"kind": "formal_proof",
"path": "submission/FixedProof.lean"
}
] | oss_fuzz | OSS-FUZZ-394126186-poc | tarantool/luajit | fixed | OSS-FUZZ-394126186-poc-formal::lean-translation-patch_validation-proof-gen | patch_validation | proof_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-122",
"CWE-787"
] | formal_task | OSS-FUZZ-394126186-poc-formal | not_gradable | the released backend has no per-profile validation commands; the released provenance lacks the patch-correspondence pointer; rebuild the instance with the current pipeline | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-394126186-poc-formal.tar.zst",
"archive_sha256": "90d9db87f149d852cfd00ff500a44d0c698b18aedd530f43aa643c02b6c0d9aa",
"archive_size_bytes": 1236499,
"checker_image_id": null,
"formal_instance_digest": "4d6343e3056bdcb132564b9a49433c1ba63e02e883478fe82e58ea952de3e325... | C | {
"declarations": [],
"files": [
{
"allowed_imports": [
"Dependency",
"SourceModel",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogramming_allowed": false,
"module": "Spec",
"path": "submission/Spec.lean"
}
],
"theorems": [
{
"al... | 20 | lean-translation-patch_validation-spec-gen | [
"source_binding",
"spec_typecheck",
"fixed_proof",
"vulnerable_disproof",
"spec_fidelity"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "specification",
"kind": "formal_specification",
"path": "submission/Spec.lean"
}
] | oss_fuzz | OSS-FUZZ-394126186-poc | tarantool/luajit | pair | OSS-FUZZ-394126186-poc-formal::lean-translation-patch_validation-spec-gen | patch_validation | spec_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-122",
"CWE-787"
] | formal_task | OSS-FUZZ-394126186-poc-formal | not_gradable | the released backend has no per-profile validation commands; the released provenance lacks the patch-correspondence pointer; rebuild the instance with the current pipeline | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-394126186-poc-formal.tar.zst",
"archive_sha256": "90d9db87f149d852cfd00ff500a44d0c698b18aedd530f43aa643c02b6c0d9aa",
"archive_size_bytes": 1236499,
"checker_image_id": null,
"formal_instance_digest": "4d6343e3056bdcb132564b9a49433c1ba63e02e883478fe82e58ea952de3e325... | C | {
"declarations": [
{
"constant": "SecurityTarget.translation",
"module": "SourceModel"
}
],
"files": [
{
"allowed_imports": [
"Dependency",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogramming_allowed": false,
"module": "SourceModel",
... | 20 | lean-translation-patch_validation-translation-gen | [
"source_binding",
"source_differential",
"translation_typecheck",
"translation_fidelity"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "translation",
"kind": "source_translation",
"path": "submission/SourceModel.lean"
}
] | oss_fuzz | OSS-FUZZ-394126186-poc | tarantool/luajit | pair | OSS-FUZZ-394126186-poc-formal::lean-translation-patch_validation-translation-gen | patch_validation | translation_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-126"
] | formal_task | OSS-FUZZ-395921923-poc-formal | legacy_lean_gate | released before per-profile validation commands; graded by the Lean gate alone (pinned obligation, kernel replay, axiom audit) | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-395921923-poc-formal.tar.zst",
"archive_sha256": "73981901955574a81ee340462c5ca5324fcdbcea03dd6a072e83fe793af6a33f",
"archive_size_bytes": 323589,
"checker_image_id": null,
"formal_instance_digest": "21934d5c491d0186e91f240dbd875bfd31fa90e79c1445faf92ec58e6e7e7b08"... | C | {
"declarations": [],
"files": [
{
"allowed_imports": [
"Dependency",
"SourceModel",
"Spec",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogramming_allowed": false,
"module": "VulnerableDisproof",
"path": "submission/VulnerableDisproof.lean... | 20 | lean-translation-patch_validation-disproof-gen | [
"source_binding",
"formal_build",
"vulnerable_disproof",
"spec_fidelity"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "vulnerable_disproof",
"kind": "checked_witness",
"path": "submission/VulnerableDisproof.lean"
}
] | oss_fuzz | OSS-FUZZ-395921923-poc | gnome/tinysparql | vulnerable | OSS-FUZZ-395921923-poc-formal::lean-translation-patch_validation-disproof-gen | patch_validation | disproof_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-126"
] | formal_task | OSS-FUZZ-395921923-poc-formal | not_gradable | the released backend has no per-profile validation commands; the released provenance lacks the patch-correspondence pointer; rebuild the instance with the current pipeline | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-395921923-poc-formal.tar.zst",
"archive_sha256": "73981901955574a81ee340462c5ca5324fcdbcea03dd6a072e83fe793af6a33f",
"archive_size_bytes": 323589,
"checker_image_id": null,
"formal_instance_digest": "21934d5c491d0186e91f240dbd875bfd31fa90e79c1445faf92ec58e6e7e7b08"... | C | {
"declarations": [
{
"constant": "SecurityTarget.translation",
"module": "SourceModel"
}
],
"files": [
{
"allowed_imports": [
"Dependency",
"Lean",
"SourceModel",
"Spec",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogram... | 20 | lean-translation-patch_validation-e2e | [
"patch_allowlist",
"source_build",
"runtime_differential",
"formal_build",
"fixed_proof",
"vulnerable_disproof",
"spec_fidelity",
"assumption_audit"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "source_patch",
"kind": "source_patch",
"path": "submission/patch.diff"
},
{
"id": "specification",
"kind": "formal_specification",
"path": "submission/Spec.lean"
},
{
"id": "vulnerable_disproof",
"kind": "checked_witness",
"path": "submission/VulnerableDisproof.l... | oss_fuzz | OSS-FUZZ-395921923-poc | gnome/tinysparql | pair | OSS-FUZZ-395921923-poc-formal::lean-translation-patch_validation-e2e | patch_validation | e2e | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-126"
] | formal_task | OSS-FUZZ-395921923-poc-formal | legacy_lean_gate | released before per-profile validation commands; graded by the Lean gate alone (pinned obligation, kernel replay, axiom audit) | {
"archive_path": "bundles/oss_fuzz/OSS-FUZZ-395921923-poc-formal.tar.zst",
"archive_sha256": "73981901955574a81ee340462c5ca5324fcdbcea03dd6a072e83fe793af6a33f",
"archive_size_bytes": 323589,
"checker_image_id": null,
"formal_instance_digest": "21934d5c491d0186e91f240dbd875bfd31fa90e79c1445faf92ec58e6e7e7b08"... | C | {
"declarations": [],
"files": [
{
"allowed_imports": [
"Dependency",
"Lean",
"SourceModel",
"Spec",
"TranslationAPI"
],
"hidden_imports": [],
"metaprogramming_allowed": true,
"module": "FixedProof",
"path": "submission/FixedProof.lean"... | 20 | lean-translation-patch_validation-proof-gen | [
"source_binding",
"formal_build",
"fixed_proof",
"assumption_audit"
] | {
"input": [
{
"content": "You are solving a formal-verification task from the llm-fv benchmark. Work only through the provided tools.\n\nWorkspace\n- Paths are relative to the workspace root. You can read every workspace file; you can write only the submission files named in the task (under `submission/`).... | 1 | [
{
"id": "fixed_proof",
"kind": "formal_proof",
"path": "submission/FixedProof.lean"
}
] | oss_fuzz | OSS-FUZZ-395921923-poc | gnome/tinysparql | fixed | OSS-FUZZ-395921923-poc-formal::lean-translation-patch_validation-proof-gen | patch_validation | proof_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-126"
] | formal_task | OSS-FUZZ-395921923-poc-formal | not_gradable | "the released backend has no per-profile validation commands; the released provenance lacks the patc(...TRUNCATED) | {"archive_path":"bundles/oss_fuzz/OSS-FUZZ-395921923-poc-formal.tar.zst","archive_sha256":"739819019(...TRUNCATED) | C | {"declarations":[],"files":[{"allowed_imports":["Dependency","SourceModel","TranslationAPI"],"hidden(...TRUNCATED) | 20 | lean-translation-patch_validation-spec-gen | [
"source_binding",
"spec_typecheck",
"fixed_proof",
"vulnerable_disproof",
"spec_fidelity"
] | {"input":[{"content":"You are solving a formal-verification task from the llm-fv benchmark. Work onl(...TRUNCATED) | 1 | [
{
"id": "specification",
"kind": "formal_specification",
"path": "submission/Spec.lean"
}
] | oss_fuzz | OSS-FUZZ-395921923-poc | gnome/tinysparql | pair | OSS-FUZZ-395921923-poc-formal::lean-translation-patch_validation-spec-gen | patch_validation | spec_gen | full | memory_vulnerability |
{
"name": "llm_fv_formal_simple_agent",
"type": "responses_api_agents"
} | lean_translation | [
"CWE-126"
] | formal_task | OSS-FUZZ-395921923-poc-formal | not_gradable | "the released backend has no per-profile validation commands; the released provenance lacks the patc(...TRUNCATED) | {"archive_path":"bundles/oss_fuzz/OSS-FUZZ-395921923-poc-formal.tar.zst","archive_sha256":"739819019(...TRUNCATED) | C | {"declarations":[{"constant":"SecurityTarget.translation","module":"SourceModel"}],"files":[{"allowe(...TRUNCATED) | 20 | lean-translation-patch_validation-translation-gen | [
"source_binding",
"source_differential",
"translation_typecheck",
"translation_fidelity"
] | {"input":[{"content":"You are solving a formal-verification task from the llm-fv benchmark. Work onl(...TRUNCATED) | 1 | [
{
"id": "translation",
"kind": "source_translation",
"path": "submission/SourceModel.lean"
}
] | oss_fuzz | OSS-FUZZ-395921923-poc | gnome/tinysparql | pair | OSS-FUZZ-395921923-poc-formal::lean-translation-patch_validation-translation-gen | patch_validation | translation_gen | full | memory_vulnerability |
End of preview. Expand in Data Studio
LLM-FV Formal Tasks
This release contains 25 task profiles derived from 5 validated formal instances. Profiles from one instance share a sealed archive; rejected backends and profiles that fail gold or slot-enforcement checks are excluded.
Task kinds: {'disproof_gen': 5, 'e2e': 5, 'proof_gen': 5, 'spec_gen': 5, 'translation_gen': 5}
Backends: {'lean_translation': 25}
Grading modes: {'legacy_lean_gate': 10, 'not_gradable': 15}
Compressed archives: 13.7 MiB
data/tasks.jsonl holds one NeMo Gym row per profile for the llm_fv_formal resources server in the llm-fv repository (integrations/nemo_gym): responses_create_params (prompt and tool schemas), agent_ref, task metadata, grading_mode (rows marked not_gradable are masked by the server), and an instance block with the release identity: the archive path, SHA-256 and size, instance and profile digests, report digests and required images. The server verifies and extracts the archive, shows the agent only the files its profile exposes, and grades in an offline sandbox. formal-tasks restore verifies the archives against these rows. data/images.jsonl records immutable Docker Hub references, original validation identities, and any archive-reassembly evidence. Generated indexes and archives are release data, not source-control inputs.
The archives include hidden reference artifacts and graders for reproducible training and local evaluation. They are not a blind held-out benchmark. Upstream source retains its original license and notices.
- Downloads last month
- 89