Add reproducible evaluation installation and weight setup

README.md

@@ -9,7 +9,7 @@ configs:
 
 # Helixbench
 
-A frozen 1,015-sample audio-video benchmark. This private staging 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.
+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 |
 |---|---:|---|
@@ -22,19 +22,30 @@ A frozen 1,015-sample audio-video benchmark. This private staging release contai
 
 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
+
+Download and unpack `evaluation.zip`. On Linux with an NVIDIA GPU, a compatible driver, `python3`, `git`, `curl`, `ffmpeg`, and `ffprobe`, run:
+
+```bash
+R=/absolute/path/to/helixbench_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](evaluation/provenance/DEPENDENCIES.md).
+
 ## Evaluate generated videos
 
-Unpack `evaluation.zip`, then install from its `evaluation` directory. 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`.
+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.
 
 ```bash
-python -m pip install -e evaluation
-helixbench verify-package
-helixbench prepare --generated /absolute/generated.jsonl --model-id my_model --run-dir /absolute/run
-helixbench run --run-dir /absolute/run --metric all --deps-root /absolute/deps --weights-root /absolute/weights --torch-home /absolute/torch
-helixbench aggregate --run-dir /absolute/run
-helixbench verify --run-dir /absolute/run
+"$R/env/bin/helixbench" prepare --generated /absolute/generated.jsonl --model-id my_model --run-dir /absolute/run
+"$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 runner needs the exact external model repositories and weights listed in `evaluation/provenance/DEPENDENCIES.md`, plus `ffmpeg` and `ffprobe`. 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.
+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. It has not been checked against every historical model end to end; the original full1015 pipeline was previously validated on an A10G with historical AudioX outputs.

RELEASE_FILES.json

@@ -1,23 +1,31 @@
 {
   "README.md": {
-    "bytes": 3301,
-    "sha256": "ba0e4043e20750ef19e05b7dcfe26b956b47d1e227ecb9b615bceefa39b9ae45"
+    "bytes": 3971,
+    "sha256": "3ec3876373f75f259132429595faa84e494a346bbbabbb28e0915bf4c7947d28"
   },
   "data/test.zip": {
     "bytes": 568170324,
     "sha256": "c6a48f1535d36f45a01cce2cb1e6eb4117319aa984c80b067ba1e89497c7d38c"
   },
   "evaluation.zip": {
-    "bytes": 6706128,
-    "sha256": "c799dc58a457a6186612f6c6549e04725060ff907d1250b6c88abe733c3ce076"
+    "bytes": 6431892,
+    "sha256": "0d6373079f273ee786244d3799f8ac9cdb0cea33a74f3996ec3ab198fff6d90a"
+  },
+  "evaluation/download_weights.py": {
+    "bytes": 3708,
+    "sha256": "02516b5d6d9f8567bc1aefe295004c3d5d5605eabb0345d6de45acc8d32d9644"
   },
   "evaluation/examples/INPUT_FORMAT.md": {
     "bytes": 573,
     "sha256": "908142d777a71d2365da374bde425db115a9db2cb944289e26974ac232fff14a"
   },
+  "evaluation/install.sh": {
+    "bytes": 2949,
+    "sha256": "6bb4f7b39e9236cea462cfaed2b0399499ede5869e09e4ca695c98182b1bdac1"
+  },
   "evaluation/provenance/DEPENDENCIES.md": {
-    "bytes": 1353,
-    "sha256": "481329fa4333cd4159a9d6e62cdbdceda17c82b3d482705c0fd5f339faa4e248"
+    "bytes": 1765,
+    "sha256": "14ba27d13793f1706efb34be949df5dc6f476a4e43b81a33b7461c492edf5117"
   },
   "evaluation/pyproject.toml": {
     "bytes": 489,

evaluation.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c799dc58a457a6186612f6c6549e04725060ff907d1250b6c88abe733c3ce076
-size 6706128
+oid sha256:0d6373079f273ee786244d3799f8ac9cdb0cea33a74f3996ec3ab198fff6d90a
+size 6431892

evaluation/download_weights.py

@@ -0,0 +1,97 @@
+"""Download and hash-check the exact public model weights used by Helixbench.
+
+Usage: /absolute/runtime/env/bin/python evaluation/download_weights.py /absolute/runtime
+"""
+from __future__ import annotations
+
+import hashlib
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+
+WEIGHTS = (
+    (
+        "weights/vggish-10086976.pth",
+        "https://github.com/harritaylor/torchvggish/releases/download/v0.1/vggish-10086976.pth",
+        "10086976245803799d9194e9a73d9b6c1549c71d1b80106f5cade5608a561f4b",
+    ),
+    (
+        "weights/synchformer_state_dict.pth",
+        "https://github.com/hkchengrex/MMAudio/releases/download/v0.1/synchformer_state_dict.pth",
+        "8aff082f2df5c3bc52759db0c865c7ee772ae6400b860d1b7e90413f2defb67c",
+    ),
+    (
+        "weights/music_speech_audioset_epoch_15_esc_89.98.pt",
+        "https://huggingface.co/lukewys/laion_clap/resolve/main/music_speech_audioset_epoch_15_esc_89.98.pt",
+        "51c68f12f9d7ea25fdaaccf741ec7f81e93ee594455410f3bca4f47f88d8e006",
+    ),
+    (
+        "deps/ImageBind/.checkpoints/imagebind_huge.pth",
+        "https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth",
+        "d6f6c22bedcc90708448d5d2fbb7b2db9c73f505dc89bd0b2e09b23af1b62157",
+    ),
+    (
+        "torch/hub/checkpoints/passt-s-f128-p16-s10-ap.476-swa.pt",
+        "https://github.com/kkoutini/PaSST/releases/download/v0.0.1-audioset/passt-s-f128-p16-s10-ap.476-swa.pt",
+        "302903fa8c4aee817b11dc982da0b29aaf8d11a3e722420476d0a12c9db70c2c",
+    ),
+)
+EVAL_DIR = Path(__file__).resolve().parent
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for block in iter(lambda: stream.read(8 * 1024 * 1024), b""):
+            digest.update(block)
+    return digest.hexdigest()
+
+
+def main() -> None:
+    if len(sys.argv) != 2 or not Path(sys.argv[1]).is_absolute():
+        raise SystemExit("Usage: python evaluation/download_weights.py /absolute/path/to/runtime")
+    prefix = Path(sys.argv[1])
+    results = []
+    for relative, url, expected in WEIGHTS:
+        target = prefix / relative
+        target.parent.mkdir(parents=True, exist_ok=True)
+        if not target.is_file() or sha256(target) != expected:
+            temporary = target.with_name(target.name + ".part")
+            subprocess.run(
+                ["curl", "-fL", "--retry", "6", "--retry-all-errors", url, "-o", str(temporary)],
+                check=True,
+            )
+            actual = sha256(temporary)
+            if actual != expected:
+                raise ValueError(f"Weight SHA256 mismatch: {relative}: {actual}")
+            temporary.replace(target)
+        results.append({"path": relative, "sha256": expected, "bytes": target.stat().st_size})
+        print(f"verified {relative}", flush=True)
+
+    expected_dino = json.loads((EVAL_DIR / "src/helixbench/resources/dino_expected.json").read_text())
+    dino = prefix / "weights/dino"
+    snapshot_download(
+        repo_id="IDEA-Research/grounding-dino-tiny",
+        revision=expected_dino["revision"],
+        allow_patterns=list(expected_dino["model_files"]),
+        local_dir=dino,
+        token=False,
+    )
+    for name, expected in expected_dino["model_files"].items():
+        if sha256(dino / name) != expected:
+            raise ValueError(f"DINO SHA256 mismatch: {name}")
+    results.append({
+        "path": "weights/dino",
+        "revision": expected_dino["revision"],
+        "files": len(expected_dino["model_files"]),
+    })
+    (prefix / "WEIGHTS_READY.json").write_text(json.dumps(results, indent=2) + "\n")
+    print("All weights ready", flush=True)
+
+
+if __name__ == "__main__":
+    main()

evaluation/install.sh

@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+# Install the exact third-party source revisions used by the six-metric runner.
+# Usage: bash evaluation/install.sh /absolute/path/to/runtime
+set -euo pipefail
+
+EVAL_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PREFIX="${1:?Usage: bash evaluation/install.sh /absolute/path/to/runtime}"
+if [[ "$PREFIX" != /* ]]; then
+  echo "The runtime path must be absolute: $PREFIX" >&2
+  exit 2
+fi
+for executable in python3 git ffmpeg ffprobe curl; do
+  command -v "$executable" >/dev/null || { echo "Missing executable: $executable" >&2; exit 2; }
+done
+mkdir -p "$PREFIX"/{deps,weights,torch}
+
+python3 -m venv "$PREFIX/bootstrap_env"
+"$PREFIX/bootstrap_env/bin/python" -m pip install 'uv==0.12.24'
+UV="$PREFIX/bootstrap_env/bin/uv"
+"$UV" python install 3.12.15
+if [[ ! -x "$PREFIX/env/bin/python" ]]; then
+  "$UV" venv --python 3.12.15 "$PREFIX/env"
+fi
+"$UV" pip install --python "$PREFIX/env/bin/python" \
+  torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 \
+  --index-url https://download.pytorch.org/whl/cu128
+"$UV" pip install --python "$PREFIX/env/bin/python" \
+  numpy==2.5.2 scipy==1.18.1 transformers==4.57.3 librosa==0.11.0 \
+  timm==1.0.28 av==16.1.0 einops==0.8.1 fvcore==0.1.5.post20221221 \
+  fairscale==0.4.13 iopath==0.1.10 ftfy==6.3.1 omegaconf==2.3.0 \
+  huggingface-hub==0.36.2 boto3 soundfile scikit-learn pandas matplotlib \
+  h5py braceexpand webdataset wget regex tqdm packaging sentencepiece \
+  opencv-python-headless imageio-ffmpeg==0.6.0 psutil torchlibrosa==0.1.0 \
+  progressbar2 resampy 'setuptools==78.1.0' 'wandb==0.30.0'
+
+while read -r name repo revision; do
+  destination="$PREFIX/deps/$name"
+  if [[ ! -d "$destination/.git" ]]; then
+    git clone "https://github.com/$repo.git" "$destination"
+  fi
+  git -C "$destination" fetch origin "$revision"
+  git -C "$destination" checkout --detach "$revision"
+  actual="$(git -C "$destination" rev-parse HEAD)"
+  if [[ "$actual" != "$revision" ]]; then
+    echo "Wrong Git revision in $destination: $actual" >&2
+    exit 1
+  fi
+done <<'REPOS'
+av-benchmark hkchengrex/av-benchmark f351b9a6fc6abde746d5f8e1d4c47c883319cb41
+ImageBind facebookresearch/ImageBind 53680b02d7e37b19b124fa37bae4b6c98c38f5be
+CLAP LAION-AI/CLAP 590b52df03a5e480634b8714135d3534313be860
+hear21passt kkoutini/passt_hear21 5f1cce6a54b88faf0abad82ed428355e7931213a
+REPOS
+
+# The benchmark's ImageBind entrypoint requires this frozen compatibility shim.
+cp "$EVAL_DIR/src/helixbench/engines/imagebind_compat_init.py" \
+  "$PREFIX/deps/ImageBind/imagebind/__init__.py"
+"$UV" pip install --python "$PREFIX/env/bin/python" --no-deps -e "$PREFIX/deps/CLAP"
+"$UV" pip install --python "$PREFIX/env/bin/python" --no-deps -e "$EVAL_DIR"
+"$PREFIX/env/bin/helixbench" verify-package
+"$UV" pip freeze --python "$PREFIX/env/bin/python" > "$PREFIX/environment.freeze.txt"
+echo "Source and Python dependencies ready in $PREFIX"
+echo "Next: $PREFIX/env/bin/python $EVAL_DIR/download_weights.py $PREFIX"

evaluation/provenance/DEPENDENCIES.md

@@ -1,14 +1,25 @@
-# 第三方依赖与权重
+# Third-party dependencies and weights
 
-本目录不内嵌第三方仓库或模型权重。运行前把仓库置于 `--deps-root` 下：
+Run `bash evaluation/install.sh /absolute/runtime` and then
+`/absolute/runtime/env/bin/python evaluation/download_weights.py /absolute/runtime`.
+These scripts fetch third-party source and weights; this release does not redistribute
+them. The source repositories must appear under `--deps-root`:
 
-| 路径 | 固定版本 | 用途 |
+| Path | Upstream repository and fixed commit | Purpose |
 |---|---|---|
-| `av-benchmark/` | `hkchengrex/av-benchmark` commit `f351b9a6fc6abde746d5f8e1d4c47c883319cb41` | VGGish、Synchformer 实现 |
-| `ImageBind/` | `facebookresearch/ImageBind` commit `53680b02d7e37b19b124fa37bae4b6c98c38f5be` | ImageBind |
-| `CLAP/src/` | `LAION-AI/CLAP` commit `590b52df03a5e480634b8714135d3534313be860` | CLAP |
-| `hear21passt/` | PaSST 源码 commit `5f1cce6a54b88faf0abad82ed428355e7931213a` | KL |
+| `av-benchmark/` | [`hkchengrex/av-benchmark`](https://github.com/hkchengrex/av-benchmark) commit `f351b9a6fc6abde746d5f8e1d4c47c883319cb41` | VGGish, Synchformer |
+| `ImageBind/` | [`facebookresearch/ImageBind`](https://github.com/facebookresearch/ImageBind) commit `53680b02d7e37b19b124fa37bae4b6c98c38f5be` | ImageBind |
+| `CLAP/src/` | [`LAION-AI/CLAP`](https://github.com/LAION-AI/CLAP) commit `590b52df03a5e480634b8714135d3534313be860` | CLAP |
+| `hear21passt/` | [`kkoutini/passt_hear21`](https://github.com/kkoutini/passt_hear21) commit `5f1cce6a54b88faf0abad82ed428355e7931213a` | KL |
 
-`--weights-root` 至少放 `vggish-10086976.pth`、`synchformer_state_dict.pth`、`music_speech_audioset_epoch_15_esc_89.98.pt` 及本地 `dino/` 模型。ImageBind 权重按其原仓库的加载约定缓存。`--torch-home/hub/checkpoints/passt-s-f128-p16-s10-ap.476-swa.pt` 的 SHA256 必须是 `302903fa8c4aee817b11dc982da0b29aaf8d11a3e722420476d0a12c9db70c2c`。PaSST 使用 CPU FP32；VGGish 使用 CPU FP32；CLAP、ImageBind、Synchformer、DINO 需要与历史环境相容的 PyTorch/CUDA。
+`download_weights.py` contains the public download URLs and full SHA256 for
+PaSST, VGGish, Synchformer, CLAP, ImageBind, and Grounding DINO. The PaSST
+checkpoint belongs at `--torch-home/hub/checkpoints/passt-s-f128-p16-s10-ap.476-swa.pt`
+and has SHA256 `302903fa8c4aee817b11dc982da0b29aaf8d11a3e722420476d0a12c9db70c2c`.
+The packaged GT PaSST and VGGish features were extracted with these models.
 
-历史 Python/CUDA 依赖版本与第三方权重来源可查原目录 `benchmark1015_all_metrics_v3/bootstrap.sh`、`extras.sh`、`weights.py`、`runtime_provenance.json`。这份整理版尚未在干净环境完成全依赖安装和 MP4 全流程验收。
+PaSST and VGGish run on CPU FP32. CLAP, ImageBind, Synchformer, and
+Grounding DINO require a compatible NVIDIA GPU. The installation script pins
+the Python and major runtime versions used in the historical A10G validation.
+The public package has not yet passed a complete fresh-environment six-metric
+run; installation and score validation are separate checks.