Phoenix-Bench

Phoenix-Bench is a collection of 511 hardware-engineering issue-resolution instances. Each instance connects an issue description, a source snapshot, a developer patch, and a testbench directory through the same instance identifier.

• Code and evaluation scripts: zanehpn/Phoenix-Bench
• Dataset files: Files and versions
• Machine-readable documentation: Croissant and Responsible AI metadata
• Small inspection sample: three sample records and issue texts and developer patches

Contents and layout

Path	Contents
issues/<instance_id>	Released issue text; some records include PR text or reproducer notes.
snapshots/<instance_id>/	Source snapshot associated with the task.
patches/<instance_id>.patch	Reference developer patch; evaluation material, not an agent input.
testbench/<instance_id>/	Per-instance verification scripts, test sources, and available logs.
testbench/_*	Shared verification helpers.
data/test.jsonl	511-row index explicitly selected by the dataset viewer.
examples/	Three small inspection examples, copied from the evaluation corpus.
f2p_full_sweep_2026-05-27.csv	Historical verification sweep; its date identifies the run, not a guarantee for later toolchain versions.

The viewer contains a lightweight index, not all repository sources. The full artifact directories remain available under Files and versions. The explicit JSONL configuration prevents unrelated CSV, source, and metadata files from being inferred as incompatible dataset splits.

Schema and split

There is one configuration, default, and one evaluation split, test, containing all 511 instances. No training or validation partition is defined. The examples are a subset of test, not additional independent evaluation instances.

Field	Type	Meaning
instance_id	string	Shared identifier connecting the four artifact locations.
issue_text	string	Full released issue text, without rewriting or truncation.
issue_url	string	Revision-pinned download URL for the original issue text.
developer_patch_url	string	Revision-pinned download URL for the reference patch.
snapshot_url	string	Revision-pinned source-directory browser URL.
testbench_url	string	Revision-pinned verification-directory browser URL.

The index follows the 511-instance record set in the published Croissant file. Issue bytes were checked against its SHA-256 values. Artifact links retain that metadata's fixed revision, 957dca0fb6d4316f0c298a5589f65795b38eeee6; adding this index does not modify task content or evaluation outcomes.

Load and inspect

from datasets import load_dataset

tasks = load_dataset("OvertureMMXXI/Phoenix-Bench", split="test")
assert len(tasks) == 511
print(tasks[0]["instance_id"])
print(tasks[0]["issue_text"])

To download just the three small inspection examples:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="OvertureMMXXI/Phoenix-Bench",
    repo_type="dataset",
    allow_patterns=["examples/**"],
    local_dir="phoenix-sample",
)

These samples contain issue descriptions, developer patches, and links to the associated source/testbench directories. They support inspecting data quality without downloading the full dataset; they are not self-contained runnable environments or a representative performance subset.

To retrieve one instance's artifacts, retain their relative directory structure:

from huggingface_hub import snapshot_download

case = "AUCOHL_Lighter_pr16"
snapshot_download(
    repo_id="OvertureMMXXI/Phoenix-Bench",
    repo_type="dataset",
    revision="957dca0fb6d4316f0c298a5589f65795b38eeee6",
    allow_patterns=[
        f"issues/{case}", f"patches/{case}.patch",
        f"snapshots/{case}/**", f"testbench/{case}/**", "testbench/_*",
    ],
    local_dir="phoenix-data",
)

Follow the code repository for toolchain setup and evaluation commands. Runner requirements vary by instance; loading the JSONL index does not execute verification. Pin both the dataset revision and code commit when reporting results. Keep reference patches and verification internals separate from the agent's editable workspace.

Collection, intended use, and limitations

Instances originate from public hardware repositories and associated issue/PR material. Use the corpus to evaluate repository-level hardware maintenance and repair under the released verification conditions. It covers Verilog/SystemVerilog and open-source EDA workflows; it does not establish performance on proprietary industrial IP, every HDL, or complete physical-design/sign-off flows.

Public-source selection and filtering can bias coverage toward particular projects and task types. Issue descriptions vary in completeness, and some include reference PR content that may expose repair information. Inspect and document prompt construction; do not assume every raw text field is leakage-free. Public repository material may also have appeared in model training data. Passing the provided checks is not proof of unrestricted hardware correctness or deployment safety.

Better evaluation can support more reliable hardware maintenance, but incorrectly accepted patches can introduce functional or security defects. Use independent review and appropriate engineering verification before deployment. Source material may retain public contributor names, usernames, emails, and attribution. Detailed collection provenance, bias, use-case, and social-impact notes are provided in the Croissant RAI fields; any synthetic-data classification there should be interpreted in conjunction with the construction of verification artifacts described in the paper.

Licensing

The published Croissant file lists MIT for Phoenix-Bench's own contribution. Redistributed repository snapshots, patches, and attribution material retain their respective upstream terms; this is not a blanket MIT relicensing of upstream sources. Inspect the license and NOTICE files associated with each source repository before reuse. A consolidated per-instance upstream-license inventory has not been added by this viewer/documentation update.

Maintenance

Report dataset access or packaging issues through the dataset discussions, and code issues through the public GitHub repository. Include the instance identifier and dataset/code revisions.