morealcholplz/ttt-vla-robomme-batch16-benchmark

TTT-VLA RoboMME — 2-GPU Batch-16 Benchmark Export

What this is

Model export from the two-GPU batch-16 throughput/memory benchmark for RoboMME. This is a benchmark artifact and should not be interpreted as a final accuracy result.

• Experiment variant: 2026-08-10 batch16_2gpu_bench
• Original local path: /home/work/mntvol/runs/ttt-vla-nuri/20260810_221822_batch16_2gpu_bench/nuri_batch16_2gpu_bench
• Repository type: model export
• Visibility: public
• Related logs/evaluation archive: https://huggingface.co/datasets/morealcholplz/ttt-vla-robomme-early-runs-eval-archive

The three benchmark/preflight repositories are intentionally separate. Their first model shard is shared, but their second shard differs; do not merge them or substitute one for another.

Contents

The repository root contains the exported model configuration and weight/tokenizer files produced by the corresponding run. The export is kept at repository root so it can be downloaded directly as a local model directory.

This artifact is preserved for reproducibility and later inspection. It does not by itself document a final RoboMME success rate.

Download

Using the Hugging Face CLI:

hf download morealcholplz/ttt-vla-robomme-batch16-benchmark --local-dir ./$(basename morealcholplz/ttt-vla-robomme-batch16-benchmark)

Using Python:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="morealcholplz/ttt-vla-robomme-batch16-benchmark",
    local_dir="./ttt-vla-robomme-batch16-benchmark",
)

Then point the TTT-VLA/RoboMME project loader at the downloaded directory. Loading is project-code specific; use the same model class and preprocessing code that produced the original export rather than assuming a generic AutoModel interface.

Provenance and caveats

• This is an artifact from the local ttt-vla-nuri experiment workspace, not a newly trained official RoboMME release.
• The exact training/evaluation logs and videos are in the related archive dataset repository.
• The base model, code, and RoboMME dataset retain their respective licenses and terms.