---
license: apache-2.0
language:
- en
tags:
- cuda
- gpu-acceleration
- code-generation
- benchmark
- llm-evaluation
- sample
size_categories:
- n<1K
configs:
- config_name: tasks
  data_files: tasks.parquet
  default: true
---

# AccelEval — Reviewer Sample (small split only)

A **representative sample** of [AccelEval](https://huggingface.co/datasets/Accel-Eval/AccelEval-data)
provided per NeurIPS 2026 dataset-hosting guidelines for reviewers
inspecting a >4 GB dataset.

> ⚠️ This dataset is uploaded **anonymously** to support a double-blind
> conference submission. Author information is intentionally omitted.

## Relationship to the full dataset

| | Sample (this repo) | Full dataset |
|---|---|---|
| Tasks | All 42 tasks | All 42 tasks |
| Scales | **`small` only** | `small`, `medium`, `large` |
| Packed size | ~30 MB | ~12 GB |
| Manifest | `tasks.parquet`, `tasks.csv` (full) | identical |
| URL | this repo | [`Accel-Eval/AccelEval-data`](https://huggingface.co/datasets/Accel-Eval/AccelEval-data) |

The two repos share an identical 42-row `tasks.parquet` (the manifest is
not affected by which scales ship), so reviewers can browse every task's
metadata, full `cpu_reference.c` source, and L1/L2/L3 `prompt_template.yaml`
directly here without needing the larger tarballs.

### How the sample was created

`small.tar.gz` is the unmodified `small`-scale tarball from the full
dataset — same SHA256, same byte layout. Specifically, every task's
`gen_data.py` is run with the parameters listed under `"input_sizes.small"`
in `tasks/<task_id>/task.json`, with the deterministic seed
(`seed=42`); the resulting `input.bin`, `expected_output.txt`,
`cpu_time_ms.txt`, and `requests.txt` files for all 42 tasks are then
packed under `tasks/<task_id>/data/small/...` into a single
gzipped tar. The same code path produces `medium.tar.gz` and
`large.tar.gz` in the full dataset; this sample simply omits those two
tarballs.

The `manifest.json` shipped here contains the SHA256 of `small.tar.gz`,
which matches the entry for `small.tar.gz` in the full dataset's
`manifest.json`.

## What's in this repo

| File | Role |
|---|---|
| `tasks.parquet` | The 42-task manifest (browseable above), 13 columns. **Includes the full `cpu_reference.c` source and the LLM `prompt_template.yaml` for every task** so the benchmark is self-contained at the metadata level. |
| `tasks.csv` | Same content as `tasks.parquet`, plain text. |
| `small.tar.gz` | All 42 tasks at the *small* scale (~30 MB packed). |
| `manifest.json` | SHA256 of `small.tar.gz` + per-task metadata. |

`small.tar.gz` expands to `tasks/<task_id>/data/small/{input.bin,
expected_output.txt, cpu_time_ms.txt, requests.txt}`.

## Quick load (manifest table)

```python
import pandas as pd
tasks = pd.read_parquet("hf://datasets/Accel-Eval/AccelEval-sample/tasks.parquet")
print(tasks.head())
```

## Pulling the small inputs

```bash
huggingface-cli download Accel-Eval/AccelEval-sample small.tar.gz \
    --repo-type dataset --local-dir .
tar xzf small.tar.gz   # populates tasks/<task>/data/small/
```

Or via the companion code repo's helper:

```bash
ACCELEVAL_HF_REPO=Accel-Eval/AccelEval-sample \
    python3 scripts/download_data.py small
```

## Categories (same as the full dataset)

| Category | Count | Examples |
|---|---:|---|
| Operations research | 12 | `crew_pairing`, `network_rm_dp`, `pdlp`, `gittins_index`, ... |
| Graph algorithms | 7 | `bellman_ford`, `held_karp_tsp`, `gapbs_pagerank_pullgs`, ... |
| Spatial--temporal | 7 | `dbscan`, `dtw_distance`, `euclidean_distance_matrix`, ... |
| HPC reference | 7 | `hpcg_spmv_27pt`, `npb_lu_ssor_structured`, `miniWeather`, ... |
| Financial computing | 5 | `black_scholes`, `bonds_pricing`, `monte_carlo`, ... |
| Scientific simulation | 4 | `sph_cell_index`, `sph_forces`, `sph_position`, `hotspot_2d` |

The full row-by-row description is in `tasks.parquet`.

## Determinism

Every task generator uses a fixed seed (`seed=42` in `gen_data.py`); the
tarball in this repo is byte-reproducible from the code repo's
`tasks/<task>/gen_data.py`. The `manifest.json` ships SHA256 hashes so
an independent regeneration can be verified bit-exactly.

## License

Released under Apache-2.0. Each individual task adapts code from a
publicly available source (GitHub repository or published research paper);
the original sources retain their respective licenses, and `tasks.parquet`
records each source for attribution.