Helixbench

A frozen 1,015-sample audio-video benchmark. This public release contains the first frame and caption_av only for each sample, plus the six-metric evaluation entrypoint. The sample IDs and prompts are fixed.

Metric	Scope	Protocol
KL ↓	1,015	PaSST 527-class softmax KL(GT‖generated), paired by ID
FAD ↓	1,015	VGGish features on the first five seconds, Fréchet distance
DeSync (s) ↓	165 event	Synchformer mean absolute offset
IB ↑	1,015	ImageBind paired video/audio cosine
CLAP ↑	1,015	LAION-CLAP generated audio against caption_av
Spatial ↑	150 perspective	v3 all-150 mean: invalid -100, eligible center 0, other clips use signed lateral-frame correctness

The dataset is data/test.zip in ImageFolder format. metadata.jsonl also exposes the AV-only captions and IDs as a directly readable file. The evaluation source is browsable under evaluation/; evaluation.zip additionally bundles the verified GT features. Each row has sample_id, caption_av, scene, interaction, and first_frame_sha256; file_name points to the PNG inside the archive. No video-only or audio-only captions are in the release. evaluation.zip contains the Python package and frozen GT PaSST/VGGish features. The Spatial target vocabulary is frozen in evaluation/src/helixbench/resources/spatial_targets150.jsonl; this preserves the historical detector prompts without releasing other captions.

Install the evaluation environment

On Linux with an NVIDIA GPU, a compatible driver, python3, git, curl, unzip, ffmpeg, and ffprobe, run:

mkdir -p helixbench && cd helixbench
curl -fL --retry 3 -o evaluation.zip https://huggingface.co/datasets/NoizAI/Helixbench/resolve/main/evaluation.zip
unzip -q evaluation.zip
R="$PWD/runtime"
bash evaluation/install.sh "$R"
"$R/env/bin/python" evaluation/download_weights.py "$R"
"$R/env/bin/helixbench" verify-package

install.sh creates a Python 3.12.15 environment with PyTorch 2.7.0/CUDA 12.8, clones four public model repositories at fixed Git commits, applies the included ImageBind compatibility shim, and installs the evaluation package. download_weights.py downloads public weights and checks their SHA256, including PaSST and Grounding DINO. Keep the unpacked evaluation directory: the editable installation and packaged GT features refer to it. See the exact dependency list.

Evaluate generated videos

Provide one MP4 for every sample ID, at 768×512, 121 frames/24 fps, stereo 48 kHz audio with at least 5 seconds. The generated JSONL rows must contain sample_id, absolute path, and SHA256 sha256. A complete manifest has exactly 1,015 rows.

"$R/env/bin/helixbench" prepare --generated /absolute/generated.jsonl --model-id my_model --run-dir /absolute/run
"$R/env/bin/helixbench" run --run-dir /absolute/run --metric all --deps-root "$R/deps" --weights-root "$R/weights" --torch-home "$R/torch"
"$R/env/bin/helixbench" aggregate --run-dir /absolute/run
"$R/env/bin/helixbench" verify --run-dir /absolute/run

The included GT features are verified by SHA256, so GT videos are not needed for KL/FAD. --gt remains available if you have the exact GT media and want to extract those features again. The entrypoint produces exactly six metrics; it does not combine them into a single score.

data/test/metadata.jsonl and all 1,015 PNG files are SHA-checked against the frozen release manifest before upload. The package's verify-package checks the frozen cohort, Spatial vocabulary and scoring engine hashes. Evaluation code is based on the private GitHub Helixbench repository at commit 1f9b199, with an AV-only data adaptation and packaged GT features. A fresh install on a single NVIDIA A10G completed all six metrics for 1,015 historical AudioX videos, followed by aggregate and verify. Compared with the prior A10G run of the same videos, KL, FAD, DeSync, IB, and CLAP matched exactly. Spatial differed by 0.0589 points because two of the 150 clips had different frame-level detector/track results; the category counts and all-150 denominator matched. Other historical models have not been rerun from this public package.