| --- |
| pretty_name: LayoutBench |
| license: cc-by-4.0 |
| language: |
| - zh |
| - en |
| tags: |
| - heap-layout-manipulation |
| - benchmark |
| - cybersecurity |
| - program-analysis |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: tasks |
| default: true |
| data_files: |
| - split: benchmark |
| path: data/tasks.jsonl |
| - config_name: references |
| data_files: |
| - split: benchmark |
| path: data/references.jsonl |
| - config_name: sample_sources |
| data_files: |
| - split: source |
| path: data/sample_sources.csv |
| --- |
| |
| # LayoutBench |
|
|
| LayoutBench is a dataset of **156 heap layout manipulation tasks**, constructed from **40 initial heap operation sequences** originating from **28 open-source projects**. It accompanies the study *LayoutScope: Difficulty-Aware Evaluation of LLM Agents for Heap Layout Manipulation*. |
|
|
| The dataset provides task descriptions, permitted compound primitives, invocation and parameter constraints, layout goals, verified reference sequences, sample source information, and frozen task attachments. The recorded platform is x86-64 Linux. |
|
|
| Version: **1.0.0**. Dataset snapshot: **2026-10-31**. |
|
|
| ## Task distribution |
|
|
| | Allocator configuration | Address reuse | Strict adjacency | Total | |
| | --- | ---: | ---: | ---: | |
| | ptmalloc 2.23 | 20 | 20 | 40 | |
| | ptmalloc 2.35 | 20 | 20 | 40 | |
| | jemalloc 5.3.1 | 18 | 18 | 36 | |
| | musl mallocng 1.2.6 | 20 | 20 | 40 | |
| | **Total** | **78** | **78** | **156** | |
|
|
| ## Files |
|
|
| | File | Contents | |
| | --- | --- | |
| | `data/tasks.jsonl` | 156 task records | |
| | `data/references.jsonl` | 156 verified reference sequences and recorded validation annotations | |
| | `data/sample_sources.csv` | 40 sample sources, covering 28 upstream projects | |
| | `attachments/A-ptmalloc-2.23.tar.gz` | Frozen public attachments for 40 tasks | |
| | `attachments/B-ptmalloc-2.35.tar.gz` | Frozen public attachments for 40 tasks | |
| | `attachments/C-jemalloc-5.3.1.tar.gz` | Frozen public attachments for 36 tasks | |
| | `attachments/D-mallocng-1.2.6.tar.gz` | Frozen public attachments for 40 tasks | |
| | `README.md` | Dataset description, field definitions, and license scope | |
| | `LICENSE` | CC BY 4.0 license text | |
| | `THIRD_PARTY_NOTICES.md` | Third-party sources and notices | |
| | `licenses/ExploitGym-DATA_LICENSE.md` | Upstream task-data licensing notice | |
| | `licenses/ExploitGym-Apache-2.0.txt` | Upstream ExploitGym source-code license text | |
|
|
| ## Data format |
|
|
| Text files use UTF-8. JSONL files contain one JSON object per line. CSV files have a header. Fields ending in `_json` contain JSON-encoded strings that can be decoded with a JSON parser. The Hugging Face configurations are `tasks`, `references`, and `sample_sources`; `tasks` is the default view. |
|
|
| Tasks and references use `task_id` as their join key. `initial_layout_id` links a task to a sample in `sample_sources`. The `benchmark` and `source` split names describe tasks and source samples. |
|
|
| ### Tasks: `data/tasks.jsonl` |
|
|
| | Field | Meaning | |
| | --- | --- | |
| | `task_id`, `task_name` | Stable frozen task identifier and human-readable name | |
| | `initial_layout_id` | Source sample identifier | |
| | `allocator_id`, `allocator` | Allocator profile identifier and display name | |
| | `relation` | Address reuse or `strict_adjacency` | |
| | `goal_id`, `goal_json` | Frozen goal identifier and logical predicate | |
| | `primitive_set_json` | Primitive definitions, parameters, object-label rules, and constraints | |
| | `runtime_json` | Recorded allocator, architecture, word size, and runtime artifact identities | |
| | `prompt`, `primitives_text` | Original task description and human-readable primitive definitions | |
| | `attachment_archive`, `attachment_directory` | Compressed archive path and task directory inside it | |
| | `target_sha256`, `target_source_sha256` | Original ELF and C source digests | |
| | `prompt_sha256`, `primitives_sha256` | Original task document digests | |
| | `task_contract`, `prompt_contract`, `primitives_contract`, `answer_contract` | Recorded task and document format versions | |
| | `target_interface`, `target_build_contract` | Recorded target interface and build format versions | |
|
|
| The decoded primitive set contains the primitive operations, parameter domains, object-label rules, per-primitive call limits where present, and total invocation constraints. Operations include allocations, deallocations, and cross-primitive frees. Goals retain their predicate version, checkpoint, constraints, and symbolic operand selectors. An operand's `epoch` identifies its selected allocation instance. |
|
|
| ### Reference sequences: `data/references.jsonl` |
|
|
| | Field | Meaning | |
| | --- | --- | |
| | `task_id`, `task_name`, `goal_id` | Task and goal identities | |
| | `reference_argv` | Ordered primitive IDs and decimal argument tokens | |
| | `reference_calls_json` | Structured calls with primitive identifiers and named arguments | |
| | `call_count` | Number of primitive invocations | |
| | `recorded_verification_runs`, `recorded_verification_hits` | Original reference validation counts | |
| | `recorded_materialization_status` | Recorded task materialization validation result | |
| | `minimization_json` | Recorded reference-sequence minimization metadata | |
|
|
| Reference sequences are published as dataset annotations. The validation fields preserve the recorded validation results associated with each reference sequence. |
|
|
| ### Sample sources: `data/sample_sources.csv` |
| |
| | Field | Meaning | |
| | --- | --- | |
| | `initial_layout_id` | Source sample identifier linked from the task table | |
| | `project_name` | Upstream project name | |
| | `upstream_entry`, `upstream_task_alias` | Upstream sample identifiers | |
| | `binary_name` | Source program binary name | |
| | `vulnerability_type` | Source vulnerability category | |
|
|
| The source heap operation sequences were collected from ExploitGym heap-vulnerability samples at the selected ASan error boundary. Project names and sample identifiers identify the data's origin. |
|
|
| ## Task attachments |
|
|
| Each archive groups tasks by allocator. A task directory is `tasks/<task_name>/public/` and contains `prompt.md`, `primitives.txt`, `target.c`, `target`, and `manifest.json`. These five files preserve the original frozen public attachments byte for byte. The task record provides the archive path and the directory to locate each task. |
|
|
| ## License |
|
|
| The dataset authors license their original contributions, including the dataset selection and arrangement, authored task descriptions, annotations, generated task specimens, source information, and documentation, under **Creative Commons Attribution 4.0 International (CC BY 4.0)**. The full license text is in `LICENSE`; the canonical license URL is https://creativecommons.org/licenses/by/4.0/. |
|
|
| This grant covers only rights that the dataset authors are entitled to grant. Third-party materials retain their applicable upstream terms. It does not relicense upstream projects, allocator libraries, ExploitGym, or other referenced third-party software. See `THIRD_PARTY_NOTICES.md` and `licenses/`. |
|
|
| The task C and ELF files are frozen specimens included with the data. Their inclusion does not expand this license to separately maintained software or files outside this dataset repository. |
|
|
| When redistributing or adapting the licensed data, provide appropriate credit, retain a link to CC BY 4.0, and indicate modifications as required by that license. The associated study title is given above. |
|
|
