latency-sensitive-bench/latency-transfer-data

Latency transfer, fixed-latency training and WanOFT timing probes — data

Training Parquet files, original evaluation records, paper plotting inputs and immutable provenance are retained here. Models have a separate experiment owner.

Sources and validation

See run_index.json, verification/native_loader_verification.json and verification/data_download_verification.json. Training files retain their original Git/LFS identities.

Five fixed-latency Flappy policies use 200 episodes (180 train/20 validation), seed 42, 2,000 steps and ordered five-frame inputs. paper/figures/fig12_wanoft_latency_grid/data.csv retains all 500 evaluation episodes across 25 conditions. The complete paper L0 source and the earlier HF L0-only evaluation remain separate evidence.

Timing probe run i4aobzgd uses the archived L2 policy and the official LIBERO step-60,000 initialization. Original reports and all available metrics are retained. The pinned run files and five output artifacts contain no classifier states. The 96 classifier states are unavailable from these sources; per-frame feature tensors were not serialized. This limitation does not remove the stored report or plotting evidence.

Experiment repository consolidation

Independent experiment sources are retained under training/<source-repository>/. Each source has its own dataset config and original metadata. Different latency units, episode budgets and action layouts are preserved. Evaluation JSON files are under evaluations/<source-model-repository>/. experiment_repository_index.json records immutable source identities and model/data/paper relationships.

Dataset contents and collection

Fifty-one Parquet configs expose the retained WanOFT context5 data, formal OpenVLA training sources and historical fixed-latency exports. Model source names and original raw dataset configs remain separate selectors. Evaluation records, paper CSVs, timing-probe reports and training code snapshots are additional archive artifacts. They are not mixed into training splits.

Fields and units

Feature schemas and split row counts come from actual Parquet metadata. Schemas are in YAML and row counts appear below. Schemas preserve encoded images, image histories, actions, prompts, episode identifiers and source latency fields. The original collector metadata defines each export's teacher, episode budget, action vocabulary, FPS and latency protocol. The formal Flappy grid uses raw frames at 30 FPS. The formal Demon grid uses 0/8/16/24/32 raw frames derived from decision settings 0/2/4/6/8 with frameskip 4 at 60 FPS. Historical latency labels retain their original semantics rather than inheriting the formal 200-episode protocol.

Source limitations

The complete paper WanOFT L0 evaluation and the earlier HF L0-only evaluation are separate files. The cumulative curriculum's declared 240-episode budget is not established by its stored 200-trajectory statistics. The original local curriculum data were not recovered. Probe classifier states and per-frame feature tensors were not found in the retained source artifacts. Shared raw Flappy L2 data have this experiment owner; scope consumers reference the same fixed revision.

Rows in each source configuration

Counts refer to stored rows, not independent episodes or unique samples across overlapping configs. New split byte sizes are not estimated for metadata fields that require decoded Arrow sizes.

Dataset config	Split	Rows
flappy_fix_latency_0_200ep_context5	train	279,882
flappy_fix_latency_0_200ep_context5	val	72,000
flappy_fix_latency_1_200ep_context5	train	297,483
flappy_fix_latency_1_200ep_context5	val	72,000
flappy_fix_latency_2_200ep_context5	train	330,734
flappy_fix_latency_2_200ep_context5	val	72,000
flappy_fix_latency_3_200ep_context5	train	316,258
flappy_fix_latency_3_200ep_context5	val	72,000
flappy_fix_latency_4_200ep_context5	train	311,588
flappy_fix_latency_4_200ep_context5	val	72,000
deadly_corridor_fix_latency_0	train	3,836
deadly_corridor_fix_latency_0	val	443
deadly_corridor_fix_latency_1	train	1,966
deadly_corridor_fix_latency_1	val	218
deadly_corridor_fix_latency_10	train	1,989
deadly_corridor_fix_latency_10	val	271
deadly_corridor_fix_latency_12	train	2,501
deadly_corridor_fix_latency_12	val	253
deadly_corridor_fix_latency_13	train	2,442
deadly_corridor_fix_latency_13	val	268
deadly_corridor_fix_latency_14	train	2,325
deadly_corridor_fix_latency_14	val	238
deadly_corridor_fix_latency_15	train	2,061
deadly_corridor_fix_latency_15	val	259
deadly_corridor_fix_latency_2	train	2,311
deadly_corridor_fix_latency_2	val	246
deadly_corridor_fix_latency_3	train	1,735
deadly_corridor_fix_latency_3	val	174
deadly_corridor_fix_latency_4	train	2,345
deadly_corridor_fix_latency_4	val	291
deadly_corridor_fix_latency_5	train	2,139
deadly_corridor_fix_latency_5	val	260
deadly_corridor_fix_latency_6	train	2,090
deadly_corridor_fix_latency_6	val	191
deadly_corridor_fix_latency_7	train	2,157
deadly_corridor_fix_latency_7	val	225
deadly_corridor_fix_latency_8	train	1,678
deadly_corridor_fix_latency_8	val	142
deadly_corridor_fix_latency_9	train	1,755
deadly_corridor_fix_latency_9	val	199
demon_attack_200ep__demon_attack_fix_latency_0_200ep	train	435,000
demon_attack_200ep__demon_attack_fix_latency_0_200ep	val	68,014
demon_attack_200ep__demon_attack_fix_latency_2_200ep	train	349,558
demon_attack_200ep__demon_attack_fix_latency_2_200ep	val	52,326
demon_attack_200ep__demon_attack_fix_latency_4_200ep	train	332,733
demon_attack_200ep__demon_attack_fix_latency_4_200ep	val	60,809
demon_attack_200ep__demon_attack_fix_latency_6_200ep	train	310,953
demon_attack_200ep__demon_attack_fix_latency_6_200ep	val	54,672
demon_attack_200ep__demon_attack_fix_latency_8_200ep	train	282,142
demon_attack_200ep__demon_attack_fix_latency_8_200ep	val	40,599
demon_attack_fix_latency_0	train	324,000
demon_attack_fix_latency_0	val	36,000
demon_attack_fix_latency_1	train	288,000
demon_attack_fix_latency_1	val	72,000
demon_attack_fix_latency_10	train	145,281
demon_attack_fix_latency_10	val	39,148
demon_attack_fix_latency_11	train	107,332
demon_attack_fix_latency_11	val	25,428
demon_attack_fix_latency_12	train	169,767
demon_attack_fix_latency_12	val	44,527
demon_attack_fix_latency_13	train	154,669
demon_attack_fix_latency_13	val	40,524
demon_attack_fix_latency_14	train	150,406
demon_attack_fix_latency_14	val	36,184
demon_attack_fix_latency_15	train	128,892
demon_attack_fix_latency_15	val	31,189
demon_attack_fix_latency_2	train	288,000
demon_attack_fix_latency_2	val	72,000
demon_attack_fix_latency_3	train	283,754
demon_attack_fix_latency_3	val	70,055
demon_attack_fix_latency_4	train	263,581
demon_attack_fix_latency_4	val	64,664
demon_attack_fix_latency_5	train	261,243
demon_attack_fix_latency_5	val	68,636
demon_attack_fix_latency_6	train	226,141
demon_attack_fix_latency_6	val	58,711
demon_attack_fix_latency_7	train	233,995
demon_attack_fix_latency_7	val	58,017
demon_attack_fix_latency_8	train	167,228
demon_attack_fix_latency_8	val	45,096
demon_attack_fix_latency_9	train	165,527
demon_attack_fix_latency_9	val	41,302
flappy_200ep__flappy_fix_latency_0_200ep	train	279,882
flappy_200ep__flappy_fix_latency_0_200ep	val	72,000
flappy_200ep__flappy_fix_latency_1_200ep	train	297,483
flappy_200ep__flappy_fix_latency_1_200ep	val	72,000
flappy_200ep__flappy_fix_latency_2_200ep	train	330,734
flappy_200ep__flappy_fix_latency_2_200ep	val	72,000
flappy_200ep__flappy_fix_latency_3_200ep	train	316,528
flappy_200ep__flappy_fix_latency_3_200ep	val	72,000
flappy_200ep__flappy_fix_latency_4_200ep	train	311,588
flappy_200ep__flappy_fix_latency_4_200ep	val	72,000
flappy_fix_latency_0_small	train	93,886
flappy_fix_latency_0_small	val	10,800
flappy_fix_latency_1_small	train	97,200
flappy_fix_latency_1_small	val	10,800
flappy_fix_latency_2_small	train	97,200
flappy_fix_latency_2_small	val	10,800
flappy_fix_latency_3_small	train	97,200
flappy_fix_latency_3_small	val	10,800
flappy_fix_latency_4_small	train	91,896
flappy_fix_latency_4_small	val	10,800

Read a named dataset configuration

This archive has 51 separate dataset configs. Choose a config explicitly rather than concatenating unrelated experiments. The revision below identifies the existing data files and selectors.

from datasets import load_dataset

rows = load_dataset(
    "latency-sensitive-bench/latency-transfer-data",
    name="flappy_200ep__flappy_fix_latency_2_200ep",
    revision="d0008f86ede35ee7db6e8d43aea75281417255b6",
    split="train",
    streaming=True,
)
episode_or_step = next(iter(rows))

Download the original files when a consumer needs the source layout:

HF_ENDPOINT=https://huggingface.co hf download latency-sensitive-bench/latency-transfer-data \
  --revision d0008f86ede35ee7db6e8d43aea75281417255b6 \
  --repo-type dataset --include 'training/flappy_200ep/flappy_fix_latency_2_200ep/**' \
  --local-dir /path/to/latency-transfer-data

Model consumers and provenance

Model archive at fixed revision records model/data relationships and identity limits. Original collection metadata and source cards remain unchanged inside the source directories. The structured indexes distinguish source facts, accepted paper inputs and unresolved bindings.

Recorded code snapshots

Available training and validation source snapshots are in the shared code archive and the fine-tuning code archive. Select the commit recorded for the run when that source is available. A compatible validation loader is not evidence of an unknown original training commit.

License and source rights

The archive has no documented common license for all retained artifacts. The YAML therefore does not assign a new archive-wide license. Original source cards, backbone terms and environment assets retain their own provenance. The license of experiment code does not establish the license of every checkpoint or rendered dataset.