Tinghong-Ye22/HFD-Bench-Eval

HFD-Bench Evaluation Core

Frozen evaluation code for the 1024-session HFD benchmark. Release: v0.1.0-hesitation-prefix-20261030-r1. This is the evaluation toolkit, not a model, a data mirror, or a completed leaderboard. It contains the same published scorer source and callable interfaces used in the author's AIStudio workspace.

给对比实验使用：先固定本仓库版本，再接入模型；不要为不同模型改指标公式、事件时间或评分参数。 Only paths, model integration settings and hardware locations may be adapted. RELEASE_LOCK.json records the exact files and fixed protocol settings; configs/path_mapping.json names relocated path fields.

Three experiment groups

Group	Input / generation	Evaluation	Release status
A: interaction	Causal real-time audio/video replay; autonomous speech and motion	Speech/motion interaction, timing, coverage, censored/missing cases	Driver and metric functions included. Speech local judge is diagnostic. Validated generated-motion semantic detection is pending.
B: quality	Same observed history, forced trigger at request completion	AnswerAcc, CER/WER, UTMOSv2; designated 14-column motion quality	Scorers and HFD reference driver included. Other models need their own controlled generation integration. Formal semantic judge is pending.
C: speech-to-motion interference	Fixed 32 pairs, with/without inserted VQA; same seed and initial state	Paired differences in the same designated motion metrics	Pairing/packing/scoring code included. Incomplete trajectories stay unavailable; surviving pairs cannot stand in for the fixed cohort.

There are three experiment groups, exported into four physical tables because speech interaction and motion interaction are separate. A GT recognizer validation score is not a model's benchmark result.

Frozen data and clocks

• Data: Tinghong-Ye22/data_vqa_all_1025, revision a6c1b2c23bf82219c7c39b4285fb9cb90d7577b5.
• Sole logical entry: releases/hesitation_prefix_20261030/metadata/final_session_delivery/.
• Pipeline: Tinghong-Ye22/data_vqa_pipeline_1025, revision ea23de1ebf62e547b04d87cec81126e38a45ce8d.
• Test: 1024 sessions. Val: 512 for calibration only. Training rows are never main-table evaluation rows.
• Audio packets: 0.6 s with 0.4 s stride; input PCM 16 kHz; output PCM 24 kHz.
• Camera: 8 Hz logical ticks over actual 5 FPS video; repeated frames are allowed, future frames are not. Motion reference: 50 Hz. Quality encoder: 30 FPS.
• Hesitation uses the actual incomplete spoken prefix. Wait until the revised question has finished; signed response latency is measured from that point. Source TTS answer onset is never the model's deadline.

Source annotations are automatically generated and not fully human reviewed. Read the metric specification for exact denominators, units, eligibility, missing observations, and source functions.

Get the fixed code snapshot

Public code repository: Tinghong-Ye22/HFD-Bench-Eval. Use tag v0.1.0-hesitation-prefix-20261030-r1, or pin its HF commit:

HF_ENDPOINT=https://hf-mirror.com hf download Tinghong-Ye22/HFD-Bench-Eval --repo-type dataset --revision v0.1.0-hesitation-prefix-20261030-r1 --local-dir HFD_bench
cd HFD_bench
chmod +x bench

This is a code snapshot; use hf download, not datasets.load_dataset.

Installation and integrity check

Use Linux/Python 3.10 in a separate evaluation environment. The tested runtime uses PyTorch 2.7.0+cu126 and Transformers 4.57.0. Install the matching PyTorch build for your CUDA host, then:

python -m pip install -r requirements-core.lock
python -m pip install -e . --no-deps
python -m hfd_bench.release --config configs/scoring.json
./bench --help

The launcher uses the active Python environment outside the native HFD checkout. Set HFD_BENCH_PYTHON to an explicit evaluation Python if needed. The native AIStudio checkout retains its existing activation behavior.

Run from the repository root. Outputs must stay under this checkout and outside source data. The supported installation is the source checkout plus editable install; a wheel alone does not contain the data, configs, or model artifacts.

Optional speech/motion dependencies are listed separately. Download public artifacts on the evaluation machine, pinned by docs/release/asset_lock.json:

python tools/fetch_locked_assets.py --list
python tools/fetch_locked_assets.py --group references
python tools/fetch_locked_assets.py --group asr
# Optional: --group judge / speech-quality / motion

The downloader uses the HF mirror with token=False. No credential is needed for these public artifacts. Third-party source archives are downloaded from their original fixed revisions, not bundled under a new license.

Run another model

1. Copy configs/scoring.json to configs/my_model.json. Edit local paths and external_adapter settings, leaving frozen fields unchanged.
2. Obtain the pinned data and a verified prepared/source input pack, or reproduce it with the HFD reference preprocessing dependencies. The source pack contains real microphone PCM, camera metadata and initial state. It is not included in this code-only repository. See reproduction.
3. Implement bench_api_version=1, identity, capabilities, reset, and run_stream as described in the adapter contract. The included external streaming plugin supports Protocol A only.
4. Run and retain raw PCM, release/playback clocks, checkpoint hash and failures:

python -m hfd_bench.release --config configs/my_model.json
./bench --config configs/my_model.json validate_gt
./bench --config configs/my_model.json run \
  --adapter python:my_adapter:MyModel --protocol A \
  --model-name MyModel --output artifacts/my_model_A
./bench --config configs/my_model.json detect \
  --run artifacts/my_model_A --calibration configs/vad_calibration.json \
  --asr-model vendor/models/whisper-large-v3 --asr-device cuda:0
./bench --config configs/my_model.json detect-inputs \
  --run artifacts/my_model_A --asr-model vendor/models/whisper-large-v3 --device cuda:0

Input ASR, output ASR, and local judging should be separate CLI invocations so each can release its GPU context. Local diagnostic judging uses judge-export, judge-local, import-intent / judge-import. Instructions and scoring commands are in reproduction.

B/C are not implemented automatically for an arbitrary plugin. An external model needs controlled/paired generation that writes the documented artifact contract. The provided HFD B/C driver is model-specific. Unsupported output branches remain N/A; A output must not be relabeled as controlled B output.

What is fixed and what remains unfinished

Code publication is frozen independently from scientific scorer calibration. This snapshot intentionally keeps detectors.frozen=false, formal semantic calibration unset, and JTS deadlines unset. --allow-unfrozen produces PROVISIONAL diagnostics, not publication-eligible main-table scores.

API clients, experimental motion recognizers, prompt sweeps, training jobs, credentials, weights, raw predictions and historical scores are excluded. Read status and validation before filling a paper table. Missing labels, unsupported branches and failed cases must stay visible.

The authors have not added a code license for this release. Public visibility does not change third-party source/model/data license terms. Dependencies and original source revisions are listed in third-party notices.