Latency transfer, fixed-latency training and WanOFT timing probes — models
The inference checkpoint bytes are unchanged from the original source. Each run retains its configuration, normalization statistics and training logs. All archived checkpoints passed the native StarVLA loader and normalization-statistics checks.
Training data and evaluation/probe evidence: latency-sensitive-bench/latency-transfer-data@23f45010.
See run_index.json for the exact source identities, code commits, W&B runs and paper consumers. Source code snapshots are in the data archive code/ directory. The original run IDs are preserved.
Consolidated experiment sources
experiment_repository_index.json records each source repository, immutable revision, retained weight identity, original training configuration and paper use. Historical trials remain separate from formal paper runs. Missing checkpoints and unpublished local data remain explicit. Original configs are provenance; current training recipes use the fixed dataset references in the index. Full provenance is in latency-sensitive-bench/latency-transfer-data@e61dcb7f75590e464444002a0951bbaca736869a.
Model details and training
| Assets | Runs or source repositories | Original training condition |
|---|---|---|
| Flappy WanOFT specialists | 5 runs | L0–L4 raw frames; ordered context5; 200 episodes; step 2000 |
| Flappy OpenVLA/QwenOFT | 7 runs | Five fixed specialists at step 5000 and two curricula |
| Demon Attack OpenVLA/QwenOFT | 5 runs | Fixed specialists at step 7000 |
| Historical fixed-latency model sources | 16 repositories | Original budgets and protocols, under historical/ |
OpenVLA here names the QwenOFT experiment framework, not a replacement for the upstream OpenVLA model family. The original backbone, initialization checkpoint, action head and normalization statistics remain in each run's configuration. The collection contains multiple policy families; it is not one merged model.
Evaluation and identity limits
The formal Flappy grid uses L0–L4 raw frames at 30 FPS and 20 episodes per cell. The formal Demon training settings 0/2/4/6/8 decision intervals correspond to 0/8/16/24/32 raw frames at 60 FPS, frameskip 4. Historical exports retain their original units and episode budgets. The paper's complete WanOFT L0 evaluation and the earlier HF L0-only file remain separate records. Timing probes use the archived WanOFT L2 policy and the recorded pre-SFT initialization.
The Flappy cumulative curriculum declares L0–L5 with 40 episodes per latency, while stored statistics report 200 trajectories and 341369 transitions. The exclusive curriculum declares L0–L4 and reports the same stored counts. The original local curriculum data were not recovered, so this card does not infer their composition. One historical Demon source is empty and has no runnable checkpoint. The 96 probe classifier states and per-frame feature tensors are unavailable from the retained sources.
Intended use and source evidence
Use this collection to reproduce the recorded simulated-control experiments. Each run is a separate artifact, with its own configuration and checkpoint identity. An archived run is a formal paper source only when its source manifest records that use.
Data and experiment evidence: latency-transfer-data at fixed revision.
Immutable run/source index records the retained identities. Original primary-run index retains the earlier archive entries.
Download and native loading
Select a run and download its whole inference bundle:
HF_ENDPOINT=https://huggingface.co hf download latency-sensitive-bench/latency-transfer-models \
--revision 8aab6023620981f9e31904e4ef4b1f11045207fb \
--include 'wan_oft_flappy_fix_latency_0_context5_2000_effbs128/**' \
--local-dir /path/to/latency-transfer-models
Keep config.yaml, config.full.yaml, normalization statistics and the selected checkpoint together.
Use the StarVLA framework and code revision recorded for that run.
Configure the recorded backbone locally and select the run-specific task wrapper.
For a LoRA bundle, retain the adapter configuration and action-head file together.
The collection root is an index rather than one AutoModel checkpoint.
Per-run metadata supplies training dates, W&B identities and code revisions when the original source recorded them. Missing source fields remain missing rather than becoming inferred training facts.
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.