video video |
|---|
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Is the v1.4.7 reasoning true to the video?
70 timelens_streaming turns for eyeballing against the clip. The text
judge that produced judge_* never saw the video, so it cannot catch reasoning
that invents visual detail and still lands on the right answer. That is what this
sample is for.
Four strata in bucket, sampled at random (seed 0):
| bucket | population | what to look for |
|---|---|---|
keep_withGT |
177,353 | the judge kept these. Does the reasoning describe what is actually on screen? |
drop_withGT |
19,495 | the judge dropped these. Was it right to? |
noGT_kept |
1,569 | count turns with no reference at all; kept on reasoning alone |
noGT_dropped |
30,626 | same, dropped on reasoning alone |
The noGT_* rows are TimeLens count questions that were never annotated;
expected_answer is the string "None" and legacy_9b holds an old Qwen3.5-9B
answer that v1.4.7 explicitly kept as provenance, not as ground truth.
Per row: question, expected (ground truth, or "None"), reasoning and
final from Qwen3.5-122B, legacy_9b, and the judge's judge_verdict /
judge_supports / judge_rationale. clip points at the mp4, cut at the fps and
frame cap the generation used (fed_fps, fed_frames), so it shows what the
model saw rather than the full-rate source.
- Downloads last month
- 93