Release readiness: working quickstart, CI, citation and community files

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,77 @@
+name: Bug report
+description: Report a spec, build, run, or verification failure
+title: "[bug] "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Build and verify failures are usually toolchain-specific, so the compiler and OS
+        versions matter as much as the error text. Please fill in every field.
+
+        Some specs are known to fail and are excluded from evaluation batches by code
+        (`EXCLUDED_SPECS` in `harness/constants.py`). Please check that list first.
+
+  - type: input
+    id: spec
+    attributes:
+      label: Spec
+      description: The spec's unique_id, or the spec file path.
+      placeholder: rodinia-nw-omp
+    validations:
+      required: true
+
+  - type: input
+    id: platform
+    attributes:
+      label: Platform and OS
+      description: CPU or GPU, distribution, and kernel or OS version.
+      placeholder: Ubuntu 24.04, x86_64, NVIDIA RTX 4070
+    validations:
+      required: true
+
+  - type: input
+    id: compiler
+    attributes:
+      label: Compiler and toolchain versions
+      description: Output of `g++ --version`, `nvcc --version`, or the OpenCL runtime version.
+      placeholder: g++ 13.2.0, nvcc 12.4
+    validations:
+      required: true
+
+  - type: textarea
+    id: command
+    attributes:
+      label: Exact command
+      description: The command you ran, verbatim, including the working directory.
+      render: shell
+      placeholder: |
+        cd ParBench
+        python3 -m harness verify specs/rodinia-nw-omp.json
+    validations:
+      required: true
+
+  - type: textarea
+    id: output
+    attributes:
+      label: Full output
+      description: The complete output, not an excerpt. Attach a file if it is very long.
+      render: shell
+    validations:
+      required: true
+
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected versus actual
+      description: What you expected to happen, and what happened instead.
+    validations:
+      required: true
+
+  - type: input
+    id: commit
+    attributes:
+      label: ParBench commit
+      description: Output of `git rev-parse --short HEAD`.
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/new-suite.yml

@@ -0,0 +1,91 @@
+name: New benchmark suite
+description: Propose adding a benchmark suite to ParBench
+title: "[suite] "
+labels: ["new-suite"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please read [docs/adding-a-suite.md](https://github.com/Scientific-Computing-Lab/ParBench/blob/main/docs/adding-a-suite.md)
+        and [CONTRIBUTING.md](https://github.com/Scientific-Computing-Lab/ParBench/blob/main/CONTRIBUTING.md)
+        before filing. Open this issue before
+        writing specs, so the pinning and licensing questions are settled first.
+
+  - type: input
+    id: suite
+    attributes:
+      label: Suite name
+      description: The slug that will prefix every unique_id, lowercase.
+      placeholder: mysuite
+    validations:
+      required: true
+
+  - type: input
+    id: upstream
+    attributes:
+      label: Upstream repository URL
+    validations:
+      required: true
+
+  - type: input
+    id: license
+    attributes:
+      label: Upstream license
+      description: ParBench does not vendor benchmark source, so the suite keeps its own license.
+      placeholder: BSD-3-Clause
+    validations:
+      required: true
+
+  - type: input
+    id: commit
+    attributes:
+      label: Pinned commit SHA
+      description: The full upstream commit the specs will be written against.
+    validations:
+      required: true
+
+  - type: checkboxes
+    id: apis
+    attributes:
+      label: APIs covered
+      options:
+        - label: CUDA
+        - label: OpenMP
+        - label: OpenCL
+        - label: OpenMP target offload
+    validations:
+      required: true
+
+  - type: textarea
+    id: kernels
+    attributes:
+      label: Proposed kernels
+      description: One line per kernel, with the API variants you intend to write specs for.
+      placeholder: |
+        stencil - CUDA, OpenMP
+        spmv - CUDA, OpenMP, OpenCL
+    validations:
+      required: true
+
+  - type: textarea
+    id: oracle
+    attributes:
+      label: Oracle plan
+      description: >-
+        For each kernel, the correctness oracle you intend to declare and its strength:
+        strong (file_hash or file_diff on a computed output), medium (numeric_comparison
+        against a tolerance), or weak (stdout_pattern plus exit_code). If a kernel can only
+        carry a weak oracle, say why - for example, the upstream references disagree across
+        APIs.
+    validations:
+      required: true
+
+  - type: textarea
+    id: build
+    attributes:
+      label: Build and run requirements
+      description: >-
+        Compilers, external data downloads, GPU requirements, and expected runtime per
+        kernel. Note anything that cannot run on a CPU-only machine.
+    validations:
+      required: false

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,21 @@
+## What this changes
+
+<!-- One or two sentences. Link the issue this closes, if there is one. -->
+
+## Checklist
+
+- [ ] `bash scripts/run_public_tests.sh` passes.
+- [ ] Every spec I added or changed validates:
+      `python3 scripts/validate_schema.py --spec specs/<unique_id>.json` exits 0.
+- [ ] Every spec I added or changed verifies against its pristine baseline:
+      `python3 -m harness verify specs/<unique_id>.json`.
+- [ ] `manifest.jsonl` changes are appended lines only. No existing entry was edited or
+      removed.
+- [ ] Run arguments were read from the source's argument parser, not from documentation
+      (`python3 scripts/spec_tools/check_spec_argc.py --all`).
+- [ ] No machine-specific paths, home directories, hostnames, credentials, or local
+      `config/paths.json` values appear in the diff.
+
+## Platform tested
+
+<!-- OS, CPU/GPU, and compiler versions the checks above were run on. -->

.github/workflows/ci.yml

@@ -0,0 +1,50 @@
+name: CI
+
+# CPU-only sanity gate for the public repo: the spec schema validates, the
+# harness package imports, and the unit suite passes on a clean clone with no
+# benchmark trees fetched and no GPU. It deliberately does NOT build or run a
+# kernel - every suite tree (rodinia, HeCBench, mixbench, xsbench, rsbench) is
+# fetched separately, see docs/benchmark-tree-pins.md.
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          # The rodinia submodule (~101 MB) buys exactly one extra passing test,
+          # and only after a local layout fixup, so CI skips it. Measured on a
+          # clean clone of the public repo: 589 passed either way.
+          submodules: false
+
+      - uses: actions/setup-python@v5
+        with:
+          # pyproject requires >=3.12; requirements-lock.txt was pinned on
+          # Ubuntu 24.04 / Python 3.12.3.
+          python-version: '3.12'
+          cache: pip
+          cache-dependency-path: requirements-lock.txt
+
+      - name: Install
+        run: |
+          python3 -m pip install --upgrade pip
+          python3 -m pip install -r requirements-lock.txt
+          python3 -m pip install -e .
+
+      - name: Validate a spec against the schema
+        run: python3 scripts/validate_schema.py --spec specs/rodinia-nw-omp.json
+
+      - name: Import the harness package
+        run: python3 -c "import harness"
+
+      - name: Unit tests
+        # The exclusion list lives in scripts/run_public_tests.sh, the same
+        # runner CONTRIBUTING.md and the PR template tell contributors to use.
+        # Measured on a clean public clone: without the exclusions the suite is
+        # 43 failed / 589 passed / 112 skipped; with them it exits 0.
+        run: bash scripts/run_public_tests.sh

CITATION.cff

@@ -0,0 +1,76 @@
+cff-version: 1.2.0
+message: "If you use ParBench, please cite both the software and the paper below."
+title: "ParBench"
+abstract: >-
+  An evaluation harness for LLM-based parallel code translation. ParBench
+  describes each benchmark kernel in a machine-readable spec and then builds,
+  runs, and verifies every LLM-produced translation, so that results are
+  comparable across models, benchmark suites, and translation directions
+  (CUDA, OpenMP, OpenCL, and OpenMP target offload).
+type: software
+version: "0.1.0"
+date-released: "2026-08-24"
+license: MIT
+repository-code: "https://github.com/Scientific-Computing-Lab/ParBench"
+url: "https://github.com/Scientific-Computing-Lab/ParBench"
+keywords:
+  - benchmark
+  - llm-evaluation
+  - parallel-computing
+  - code-translation
+  - cuda
+  - openmp
+  - opencl
+authors:
+  - given-names: "Samyak"
+    family-names: "Jhaveri"
+    affiliation: "University of California, Irvine"
+  - given-names: "Erel"
+    family-names: "Kaplan"
+    affiliation: "Technion"
+  - given-names: "Tom"
+    family-names: "Yotam"
+    affiliation: "Code Metal"
+  - given-names: "Le"
+    family-names: "Chen"
+    affiliation: "Argonne National Laboratory"
+  - given-names: "Tomer"
+    family-names: "Bitan"
+    affiliation: "Technion"
+  - given-names: "Niranjan"
+    family-names: "Hasabnis"
+    affiliation: "Code Metal"
+  - given-names: "Gal"
+    family-names: "Oren"
+    affiliation: "Stanford University, Technion"
+preferred-citation:
+  type: conference-paper
+  title: >-
+    ParBench: A Benchmark for Reliable Evaluation of LLM Parallel Code
+    Translation
+  year: 2026
+  collection-title: "Proceedings of the Conference on Neural Information Processing Systems (NeurIPS 2026)"
+  conference:
+    name: "Conference on Neural Information Processing Systems (NeurIPS 2026)"
+  authors:
+    - given-names: "Samyak"
+      family-names: "Jhaveri"
+      affiliation: "University of California, Irvine"
+    - given-names: "Erel"
+      family-names: "Kaplan"
+      affiliation: "Technion"
+    - given-names: "Tom"
+      family-names: "Yotam"
+      affiliation: "Code Metal"
+    - given-names: "Le"
+      family-names: "Chen"
+      affiliation: "Argonne National Laboratory"
+    - given-names: "Tomer"
+      family-names: "Bitan"
+      affiliation: "Technion"
+    - given-names: "Niranjan"
+      family-names: "Hasabnis"
+      affiliation: "Code Metal"
+    - given-names: "Gal"
+      family-names: "Oren"
+      affiliation: "Stanford University, Technion"