gavinlaw/multimodal-understanding-controlled-probes

Controlled multimodal understanding probes

人工绘制、程序化生成的诊断数据；用于区分观察不足、局部坐标转换失败与音画对应错误。

There are 7,496 evaluation records, including repeated controls and paired worlds; these are not 7,496 independent scenes. Temporal sensitivity controls hold the active mouth open across the speech interval. Stimuli were regenerated after the compute workspace disappeared on lease release; validation compares every model input hash against the preserved original predictions. environment.txt is reconstructed from the same base image and pinned packages, not a live snapshot of the original run.

A: 200 paired shutter-inspection scenes, four simple properties. B: 100 sampled paired map scenes from 16 cardinal rotation configurations. C: 60 paired cartoon scenes with synthetic eSpeak-NG speech; audio content stays fixed when the active character is swapped. These are narrow controlled probes, not natural-video benchmarks or proof of a new method.

stimuli.tar.gz contains all images, 6-second videos, generated speech and records.jsonl. outputs.tar.gz contains raw greedy model predictions, token accounting, environment and summaries. source/ contains generation, inference and analysis code. See manifest.json for immutable model and code revisions and SHA256 hashes.

Generate with python source/data.py --root /data/outputs/stimuli, then python source/controls.py --root /data/outputs/stimuli --extended and python source/temporal.py --root /data/outputs/stimuli (Pillow, scipy, NumPy, eSpeak-NG, ffmpeg and DejaVu Sans required). Extract stimuli under /data/outputs/ to reuse the recorded absolute asset paths. Inference uses PyTorch 2.9.1, transformers 4.57.1, qwen-omni-utils 0.0.8, greedy decoding; C uses matching 4 fps sampling and processor clocks. A/B evaluate the patch-sized Conv3d projection as an equivalent linear projection; comparison against the original operation is recorded in run logs. Attention uses the original transformers implementation. No model weights are included.

Limitations: B places the red and green objects on opposite sides of the shared landmark, so a single view can determine the joint answer; it does not establish that both sources are necessary. C has only two visual speaker-order templates and two distinct single-frame control images. Repeated records must not be treated as independent visual scenes.

To rerun inference, create an empty output directory and launch python source/worker.py --root /data/outputs/stimuli --output /data/rerun/part-00.jsonl --shard 0 with an explicitly authorized CUDA_VISIBLE_DEVICES. Repeat for shards 1, 2 and 3 using separate output files; shards 0/1 select A, 2 selects B and 3 selects C. Then run python source/analyze.py --input /data/rerun --output /data/rerun/summary.json. Existing prediction files resume by record ID. python source/validate.py --root /data/outputs checks the frozen release, including its original exit receipts; it is not a generic validator for new runs.

Raw outputs use model licenses as applicable; original procedural scenes and metadata are released under CC BY 4.0. No real people, voices, or third-party video datasets are included. C speech is synthesized by eSpeak-NG; it is not human speech. Results must not be interpreted as real-world performance, RL training gains, or a novelty claim.

Companion research code and reports: https://github.com/luojiaxuan/cvpr-multimodal-understanding (private project repository).

Iteration two

iteration-two/ adds pixel-rule observation, four-world spatial composition, timestamped audio-visual controls and a frozen 100-video TempCompass order subset. Third-party video bytes are not redistributed. Original synthetic assets retain CC BY 4.0; TempCompass questions and annotations retain CC BY-NC 4.0 with attribution and license in iteration-two/. See its README and manifest for schema, sources, model/code revisions and checksums.