offline/qwen35/
offline/mimo25/
offline/kimi26/
offline/gemini35/
offline/qwen35-omni/
streaming/aura/
streaming/minicpmo/
streaming/vst/
streaming/streamforest/
streaming/thinkstream/
streammind-v1/qwen35/
streammind-v2/qwen35/full.tar.gz
streammind-v2/gemini35/full.tar.gz
viewer/
preview/
scripts/build_hf_preview.py
Files under offline/, streaming/ and streammind-v1/ retain their original records. StreamMind v2 archives include Debug, ASR, memory databases, evidence frames and saved Judge results under judged/.
The Hugging Face Dataset Viewer uses preview/ for question-level answers and saved Judge scores; source_file and source_line locate each record. For archive sources, :: separates the archive path from the member path relative to the extracted run root. Missing scores stay null. Full StreamMind v2 replay is available through the local viewer below.
elapsed_sec is the recorded wall-clock duration, not a video timestamp. For proactive tasks it includes waiting for the event. answer_time_sec is the annotated expected event time for proactive tasks, not response latency. Missing timings remain null.
Python 3.10+ is required. The viewer itself has no third-party dependencies.
python -m pip install -U huggingface_hub
hf download hkuzxc/StreamArena_Trajectory --repo-type dataset \
--include 'viewer/*' --include 'streammind-v2/gemini35/*' --local-dir trajectories
mkdir -p runs/gemini35
tar -xzf trajectories/streammind-v2/gemini35/full.tar.gz \
-C runs/gemini35 --strip-components=1
python trajectories/viewer/review.py --run-dir runs/gemini35
Open http://127.0.0.1:8765. Select a video and question to inspect its recorded answer, worker trace and question-time Memory snapshot. The ASR / Debug / SQL link opens the complete archive inspector. Only one run is displayed.
For the Qwen run, replace gemini35 with qwen35 in the commands above.
Download a corresponding original video from StreamArena:
hf download hkuzxc/StreamArena videos/2v8ut2nqZJY.tar \
--repo-type dataset --local-dir source-videos
mkdir -p videos
tar -xf source-videos/videos/2v8ut2nqZJY.tar -C videos
python trajectories/viewer/review.py --run-dir runs/gemini35 --video-dir videos
The viewer accepts videos/<video_id>/<video_id>.mp4 or videos/<video_id>.mp4. Click a recorded timestamp to seek the video. It reads local files only and does not call any model APIs. Install FFmpeg to reconstruct video frames at recorded timestamps; reconstructed frames are labeled separately from archived Memory evidence.
For SQL inspection, select the archived session database and table. Keep any accompanying -wal and -shm files beside the database. Memory and SQL show the archived final state; question-time evidence is in the question trace.
Video rights remain with their original owners; see the StreamArena repository for the original release terms.
offline/qwen35/
offline/mimo25/
offline/kimi26/
offline/gemini35/
offline/qwen35-omni/
streaming/aura/
streaming/minicpmo/
streaming/vst/
streaming/streamforest/
streaming/thinkstream/
streammind-v1/qwen35/
streammind-v2/qwen35/full.tar.gz
streammind-v2/gemini35/full.tar.gz
viewer/
preview/
scripts/build_hf_preview.py
Files under offline/, streaming/ and streammind-v1/ retain their original records. StreamMind v2 archives include Debug, ASR, memory databases, evidence frames and saved Judge results under judged/.
The Hugging Face Dataset Viewer uses preview/ for question-level answers and saved Judge scores; source_file and source_line locate each record. For archive sources, :: separates the archive path from the member path relative to the extracted run root. Missing scores stay null. Full StreamMind v2 replay is available through the local viewer below.
elapsed_sec is the recorded wall-clock duration, not a video timestamp. For proactive tasks it includes waiting for the event. answer_time_sec is the annotated expected event time for proactive tasks, not response latency. Missing timings remain null.
Python 3.10+ is required. The viewer itself has no third-party dependencies.
python -m pip install -U huggingface_hub
hf download hkuzxc/StreamArena_Trajectory --repo-type dataset \
--include 'viewer/*' --include 'streammind-v2/gemini35/*' --local-dir trajectories
mkdir -p runs/gemini35
tar -xzf trajectories/streammind-v2/gemini35/full.tar.gz \
-C runs/gemini35 --strip-components=1
python trajectories/viewer/review.py --run-dir runs/gemini35
Open http://127.0.0.1:8765. Select a video and question to inspect its recorded answer, worker trace and question-time Memory snapshot. The ASR / Debug / SQL link opens the complete archive inspector. Only one run is displayed.
For the Qwen run, replace gemini35 with qwen35 in the commands above.
Download a corresponding original video from StreamArena:
hf download hkuzxc/StreamArena videos/2v8ut2nqZJY.tar \
--repo-type dataset --local-dir source-videos
mkdir -p videos
tar -xf source-videos/videos/2v8ut2nqZJY.tar -C videos
python trajectories/viewer/review.py --run-dir runs/gemini35 --video-dir videos
The viewer accepts videos/<video_id>/<video_id>.mp4 or videos/<video_id>.mp4. Click a recorded timestamp to seek the video. It reads local files only and does not call any model APIs. Install FFmpeg to reconstruct video frames at recorded timestamps; reconstructed frames are labeled separately from archived Memory evidence.
For SQL inspection, select the archived session database and table. Keep any accompanying -wal and -shm files beside the database. Memory and SQL show the archived final state; question-time evidence is in the question trace.
Video rights remain with their original owners; see the StreamArena repository for the original release terms.