vgandhi13/glimpse-video-holmes-train

Glimpse agentic trajectories on Video-Holmes (train)

Multi-turn, tool-using trajectories from the Glimpse model (elorian/glimpse, via OpenRouter) answering every question of the Video-Holmes train split (1,551 multiple-choice questions over 233 short films). The model starts with no frames. It repeatedly calls a frame_select tool to sample frames from chosen time ranges, reasons over them, and gives an answer. Every run includes the model's reasoning, its tool calls, the exact frames it received, and whether its final answer was correct.

Questions	1,551 (all of the Video-Holmes train split)
Runs	2,254 (1,277 correct, 977 incorrect)
Questions with ≥1 correct run	1,276 (82.3%)
Frames	130k PNGs (25 GB), in tar shards

Layout

README.md
index.json                              per question: runs, correct_runs, incorrect_runs, cost
solved_ids.txt                          Question IDs with at least one correct run
correct/q<ID>/run_<i>/trajectory.json   runs whose final answer matched the label
correct/q<ID>/run_<i>/raw_trace.jsonl
incorrect/q<ID>/run_<i>/...             completed runs with a wrong answer
frames/correct-<NNNNN>.tar              frame PNGs for correct runs
frames/incorrect-<NNNNN>.tar            frame PNGs for incorrect runs

• q<ID> is the zero-padded Video-Holmes Question ID (q0001 … q1551).
• run_<i> is the question's i-th attempt, numbered across both splits. A question solved on a later attempt appears in both splits, e.g. incorrect/q0012/run_1 and correct/q0012/run_2.
• Frames are packed into tar shards to keep the file count low. The tar members are q<ID>/run_<i>/frames/turn_<t>/call_<c>/frame_<k>.png. Extract a split's shards inside that split's folder, and every frames[].file path in trajectory.json resolves:

huggingface-cli download vgandhi13/glimpse-video-holmes-train --repo-type dataset --local-dir glimpse
cd glimpse
for s in correct incorrect; do for t in frames/$s-*.tar; do tar -xf "$t" -C "$s"; done; done

trajectory.json

Field	Meaning
question_id, video_id, question_type	Video-Holmes identifiers (question_type: MHR, IMC, TCI, SR, CTI, TA, PAR, …)
prompt	the exact initial messages sent to the model
gt, answer, correct	ground-truth label, model answer, whether they match
stop_reason	answer_found, no_answer_no_tool_call, max_turns_exceeded or tool_error_retry_limit
num_rounds, tool_calls, tool_errors	counted turns, total frame_select calls, failed calls
usage_total	summed prompt/completion tokens and cost (USD)
model, temperature, max_turns, max_visual_tokens	settings used for this run
turns[]	one entry per model response: reasoning, content (visible text), tool_calls[] (arguments, status, error, effective arguments after the visual-budget fallback, and frames[] with file and timestamp), finish_reason, usage

raw_trace.jsonl holds the unmodified API response for each turn.

How the trajectories were generated

Setting	Value
Model	elorian/glimpse via OpenRouter, reasoning enabled, full message and reasoning history replayed each turn
Protocol	EfficientVideoAgent (EVA)-style. One native tool, frame_select(start_time "MM:SS", end_time, nframes, resize), samples frames uniformly from an interval. Its results are returned as images with timestamps.
Prompt	EVA's system prompt plus guidance for the native tool. The user turn gives video length, resolution, the question and options A–F.
Turn budget	6 counted model responses. A response in which every tool call failed doesn't count (up to 10 extra). On the last turn the model is told to answer.
Visual budget	about 12,000 estimated visual tokens per tool call (EVA fallback: downscale to 720p, then reduce frames and resolution)
Output budget	65,536 tokens per response
Scoring	exact match of the <answer>X</answer> label against the ground-truth label
Videos	videos_cropped/ from the official Video-Holmes release

Collection ran in rounds, aiming for at least one correct trajectory per question:

Round	Questions attempted	Temperature	Correct runs	Solved after round
First pass	all 1,551	0.8	1,122 / 1,534 (73%)	1,121 (72.3%)
Retry 1	430 unsolved	0.8	117 / 416 (28%)	1,238 (79.8%)
Retry 2	313 unsolved	1.0	38 / 304 (12.5%)	1,276 (82.3%)

Each run's temperature field records which setting it used.

Caveats

• A correct answer doesn't mean correct reasoning. Each question has 6 options, so some correct runs (especially those found only on a retry) may be lucky guesses. Before training on these, you may want to filter, e.g. by checking the reasoning against the Explanation field of the Video-Holmes train annotations.
• Unsolved questions. 275 questions have no correct run after 3 attempts. Retry 2's solve rate fell below chance, which suggests Glimpse consistently picks the same wrong answer on them.
• File paths in tool results. The frame_select results returned to the model included each frame's local file path (e.g. .../sample_0134/media/<date>/<video_id>/...), and in a few runs the model's reasoning comments on these paths. In one case it reads the YouTube video ID from them. A local workspace name was replaced with [redacted] in 10 runs.
• Missing attempts. 41 attempts exceeded the model's 262,144-token context limit (many frames plus the 65,536-token output reserve) and aren't included, so some questions have fewer runs.

License and terms

The frames are derived from the Video-Holmes benchmark, which is released under Apache-2.0 for academic purposes only. All videos come from the Internet, and copyright remains with the original owners. This dataset follows the same terms: use it for academic research only. If any content infringes your rights, please open a discussion on this repository and it will be removed. Model outputs were generated with Glimpse via OpenRouter; use of them is also subject to the model provider's terms.

Please also cite Video-Holmes:

@article{cheng2025video,
  title={Video-Holmes: Can MLLM Think Like Holmes for Complex Video Reasoning?},
  author={Cheng, Junhao and Ge, Yuying and Wang, Teng and Ge, Yixiao and Liao, Jing and Shan, Ying},
  journal={arXiv preprint arXiv:2505.21374},
  year={2025}
}