DVC_F1/DVC_llm score exactly 0.000 despite submissions parsing correctly via your own process_raw_output()
Three of our submissions (qwen3vl_2B, rank 137; qwen3vl_8B, rank 128; qwen3vl_32B, rank 131; all dated 2026-09-23) scored exactly DVC_F1=0.000 and DVC_llm=0.000 on every dense_captioning_* sample, despite VS_llm (also LLM-judged) scoring normally and at expected magnitude for all three; so the LLM-judge stage clearly ran in general.
We verified this isn't a formatting issue on our end before reporting: I fetched evaluation/eval_dvc.py's process_raw_output() function directly and ran it against every one of our submitted dense_captioning_gpt/dense_captioning_gemini predictions.
100% (1,341/1,341 checked) parsed into valid {start, end, caption} segments using your own regex, with no errors. We also confirmed our submission text is structurally identical in format to an earlier submission of ours (videochat_flash, rank 124) that did score nonzero on both metrics (DVC_F1=0.210, DVC_llm=2.125) had the same pattern: "{start}-{end} seconds: {caption}", newline-separated.
Since parsing succeeds but the score is still exactly zero, we suspect the issue is downstream of parsing, possibly in group_records_by_dataset()'s dataset-bucketing heuristics (which infer dataset from video_id/data_source patterns), or in how these three specific submissions merged with private ground truth. We don't have visibility into that part of the pipeline to debug further ourselves.
Happy to share our full leaderboard_submission.json files for these three submissions if useful for reproducing this.