AVTime public code release
This directory is a self-contained, path-sanitized copy of the current AVTime data, training, and benchmark code. It contains code and frozen split IDs only; videos, annotations, model weights, generated predictions, and credentials are not bundled.
Mirrors: Hugging Face (public gated access) and GitHub (public source mirror).
| Part | Entry points | Output |
|---|---|---|
| Data | data/split_avtime_50k.py, data/caption_pipeline/ |
AVTime-50K split and merged dense captions |
| Training | training/sft/, training/bitsc/ |
SFT actor and pure BiTSC GRPO checkpoints |
| Benchmark | benchmark/run_avtime_bench.sh |
canonical target predictions, official merge, Qwen3.6-27B judgments, 3x2 scores |
Data
The default splitter reproduces the frozen 300-video zero-shot AVTime-Bench
slice from the 49,977-row AVTime corpus, yielding 49,677 training rows. See
data/README.md for the split and caption pipeline.
Training
The actor learns one-round percentage-timestamp event captions. The GRPO
dataset and ORM are pure BiTSC: time-to-event plus event-to-time only. The
release contains no SC/TVG data generator or task-specific reward branch. See
training/README.md.
AVTime-Bench
The default entry point follows the current HQ300 protocol: historical
LongVALE timestamp-caption prompting, deterministic official adjacent-caption
merge, and then AVTime scoring. Qwen3.6-27B is configured as a caption-only
judge/selector, while deterministic code owns every temporal calculation. A
direct structured-JSON inference variant is retained separately. See
benchmark/README.md.
Validate the code-only package
python3 -m pip install -r requirements-dev.txt
python3 -m pytest -q
python3 scripts/check_release.py
GPU execution additionally requires the matching PyTorch, vLLM, Qwen-Omni utilities, and Megatron-SWIFT stacks. Those fast-moving packages should be installed for the exact checkpoint/GPU platform rather than from an untested universal lockfile.
Before publishing, complete RELEASE_CHECKLIST.md,
especially dataset/weight access, third-party license review, and the 27B judge
rerun. Repository-level provenance is recorded in PROVENANCE.md.