EchoMinkki/Orbdivination

Thyme Video RL — pass-rate filtered subset

RL training data for a video agent that first skims 64 uniformly sampled frames, then calls a temporal-crop tool to pull higher-FPS segments from the source mp4.

Each row therefore carries both:

• videos — the 64 overview frames, embedded as JPEG bytes
• video_path — the source mp4, which the crop tool reads at rollout time

Provenance

Built from 12,000 multiple-choice questions over 1,479 videos (LVBench, LongVideoBench, Video-MME). Every question was rolled out 8 times with the SFT checkpoint that RL starts from, under the video_crop_v8 agent harness, then filtered by pass rate.

Stage	Questions
Source	12,000
Dropped: solved 8/8 (zero gradient)	2,495
Dropped: solved 0/8 (zero gradient)	476
Dropped: answered without tools in ≥5/8 rollouts	3,645
Kept	5,384

The pass-rate filter is DAPO's dynamic sampling moved offline: groups whose rollouts all agree carry no advantage signal. The tool-use filter additionally removes questions the model answers from the overview alone, which teach little about temporal exploration.

⚠️ Pass rates are relative to one policy. They were measured with the RL starting checkpoint at temperature=1.0; as training moves the policy, the labels drift. Treat them as a snapshot, not an intrinsic difficulty.

Schema

Mirrors the project's existing image RL data, plus video_path.

Field	Type	Notes
data_source	string	thyme_video_rl
prompt	list<{role, content}>	single user turn, contains <video>
videos	list<{bytes, path}>	64 JPEG frames, 480×270, path empty
video_path	string	videos/<name>.mp4, repo-relative
ability	string	video_qa
env_name	string	""
reward_model	{ground_truth, style}	style="rule", ground truth is the option letter
extra_info	struct	see below

extra_info: answer, index, question, split, plus the prefilter record — pass_rate, pass_count, n_rollout, source_bench, zero_tool_rollouts, video.

Resolution: read this before training

The overview frames are 270-short-side, 128 visual tokens per frame, fixed by the harness protocol that the SFT checkpoint was trained under.

This is deliberately not the resolution path used for single-image RL data. With patch16 + spatial_merge2 (one token per 32×32 pixels), a multi-megapixel image consumes up to ~2048 tokens. Applying that budget per frame here:

64 frames × 1024 tokens  = 65,536 tokens   → 2× the 32k context
192 frames (with crops)  = 196,608 tokens  → 6× the context

Configure the video path at 128 tokens/frame. Do not reuse the image path's per-image token cap.

Shard ordering

Rows follow source-benchmark order, so a single 200-row shard tends to hold one benchmark. Shuffle globally when training; reading shards in order yields benchmark-homogeneous batches.

Videos: not distributed here

The 64 overview frames are embedded, so the dataset is self-contained for the skim step. The source mp4 files are not included — they come from the public LVBench, LongVideoBench and Video-MME releases (1,445 files, ~254 GB).

video_path is a relative path in the form videos/<name>.mp4. Nothing in verl resolves this field; it exists for the temporal-crop tool, so point it at wherever you keep the videos:

VIDEO_ROOT = "/your/path/to/videos"           # holds the mp4 files directly
path = os.path.join(VIDEO_ROOT, row["video_path"].removeprefix("videos/"))

# or keep the prefix and mirror the layout:
path = os.path.join(SOME_ROOT, row["video_path"])   # <SOME_ROOT>/videos/<name>.mp4

extra_info.video carries the bare filename if you would rather build the path yourself.

videos_manifest.tsv lists every referenced video with its size, a SHA-256 of the first MiB, and how many questions use it — enough to check your local copy matches:

filename	size_bytes	sha256_first16	questions	source_bench

Videos are shared across questions (median 4 questions per video, max 11), so 1,445 files cover all 5,384 rows.

Known gaps

• 28 of the 12,000 source questions have 6–7 scored rollouts instead of 8 (rule-based answer extraction failed on one rollout). They are excluded from this subset only if the filter drops them; n_rollout records the actual denominator per row.
• Frame extraction used CFR timestamps verified against decoded PTS. Videos with variable frame rate were rejected upstream, so all 1,479 are CFR.