Refresh README.md: 227658 records

README.md

@@ -21,21 +21,19 @@ in a system prompt, because that is what the evaluation harnesses send.
 
 | | |
 |---|---|
-| Records | 179,559 |
+| Records | 227,658 |
 | Distinct videos | 19,805 |
-| Reasoning style | 107,457 |
-| Direct style | 72,102 |
+| Reasoning style | 136,496 |
+| Direct style | 91,162 |
 
 | Source dataset | Records |
 |---|---|
-| TimeLens-100K | 170,720 |
+| TimeLens-100K | 218,819 |
 | THUMOS-14 | 8,839 |
 
-TimeLens records by video source: cosmo_cap 62,013, didemo 40,561,
-internvid_vtime 36,528, queryd 18,142, hirest 13,476.
+TimeLens records by video source: cosmo_cap 79,024, internvid_vtime 53,400, didemo 48,573, queryd 22,892, hirest 14,930.
 
-Question types among reasoning records: `yes_no_count` 38,772,
-`direct_action` 34,350, `multiple_choice` 34,335.
+Question types among reasoning records: `yes_no_count` 48,452, `direct_action` 44,028, `multiple_choice` 44,016.
 
 ## Record format
 
@@ -72,11 +70,12 @@ The instruction that selects a style lives in the user turn:
   option letter.
 
 The signals are mutually exclusive across the whole file, verified:
-107,457/107,457 reasoning records carry the reasoning instruction and none carry
-the direct one; 72,102/72,102 direct records carry the direct instruction and
-none carry the reasoning one. All 107,457 reasoning targets end with
-`FINAL ANSWER:`; all 72,102 direct targets are a single line; all 34,290
-multiple-choice direct targets are one letter.
+136,496/136,496 reasoning records carry the reasoning instruction and
+0 carry the direct one; 91,162/91,162 direct records carry the
+direct instruction and 0 carry the reasoning one. All
+136,496 reasoning targets contain `FINAL ANSWER:`; all
+91,162 direct targets are a single line; all
+43,946 multiple-choice direct targets are one letter.
 
 `system` is `"You are a helpful assistant."` on every record, matching the
 default in OVO-Bench and StreamingBench evaluation. No task instruction is
@@ -93,27 +92,35 @@ Direct answers are normalised to a bare value; samples where the model hedged
 
 ## Video coverage is uneven across the run
 
-Questions per video, counting reasoning records for TimeLens:
+TimeLens has 15 questions per video: 5 events, each asked three ways. Questions
+per video actually taken, counting reasoning records:
 
 | Questions taken | Videos |
 |---|---|
-| 3 | 15,208 |
-| 6 | 517 |
+| 3 | 5,530 |
+| 5 | 1 |
+| 6 | 10,198 |
+| 8 | 1 |
 | 9 | 59 |
+| 10 | 1 |
+| 11 | 2 |
 | 12 | 187 |
-| 15 | 3,424 |
-
-The 3,424 videos with all 15 questions come from the earliest batches, which had
-no per-video cap. The cap was added later but did not bind until a bug was fixed:
-it was applied to the unprocessed questions rather than to all of a video's
-questions, so successive passes kept drilling deeper into the same videos. Once
-fixed, the bulk of the run took exactly one event (three questions) per video —
-that is the 15,208 group. Treat the 15-question videos as a distinct, denser
-subset if per-video balance matters for your training mix.
+| 13 | 3 |
+| 14 | 4 |
+| 15 | 3,425 |
+
+All 19,411 videos in TimeLens-100K are represented. The largest group is
+10,198 videos at 6 questions. Videos at 15 come from the earliest batches, which
+had no per-video cap; the cap was added later but did not bind until a bug was
+fixed, because it was applied to the unprocessed questions rather than to all of
+a video's questions. Later runs took whole events per video — 3, then 6, then 9
+questions — so the deeper groups are events 2 and 3 of videos already covered,
+not repeats of the same event. Treat the 15-question videos as a distinct,
+denser subset if per-video balance matters for your training mix.
 
 ## Known limitations
 
-- **Reasoning is short.** A median of 133 characters (p90 212). Most records are
+- **Reasoning is short.** A median of 133 characters (p90 209). Most records are
   a single grounded observation rather than a multi-step chain. This follows from
   the low thinking level and from batching three questions per request.
 - **THUMOS-14 label noise.** Reference answers were derived faithfully from
@@ -123,7 +130,7 @@ subset if per-video balance matters for your training mix.
   and did not drive generation.
 - **Not verified against ground truth.** These are model outputs, not human
   labels. Agreement with THUMOS-14 references is 50.2%; with TimeLens
-  multiple-choice references, 91.9%.
+  multiple-choice references, 95.5%.
 - **OVO-Bench overlap excluded.** The 105 THUMOS-14 test questions whose videos
   appear in OVO-Bench's REC split are excluded, since OVO-Bench uses a
   byte-identical prompt template and is an evaluation target.