| # LingBot-VA Attention Analysis Dataset |
|
|
| Attention-analysis dataset generated on `h100-server` for the RoboTwin task `grab-the-medium-sized-white-mug-rotate-it-place-it-on-the-table-and-hook-it-onto-the-smooth-dark-gray-rack`, using the LingBot-VA 1.0-style velocity FlowMatch inference path (checkpoint `lingbot-va-posttrain-robotwin`). |
|
|
| ## Content |
|
|
| - `artifacts/archives/lingbot-va-attn-trajectory-6steps.tar.gz.part00` … `part09` — the full six-step raw attention capture: 4,320 dense CSV matrices (6 steps `0,4,9,14,19,24` × 30 layers × 24 heads) plus `summary.json`, packed as a single gzip tarball (~77.7 GB) and split into ~8 GiB parts to stay within the Hub per-file limit. Archive sha256: `672366873edd322eed50879226487042f38cbd3afbe7186ba584ef565b806c39`. |
| - `derived/density/top-p-0.9/trajectory-6steps/` — per-layer top-p=0.9 attention-density CSVs, one per step, produced by `script/compute_attention_density.py`. |
| - `visualizations/attention-heatmaps/trajectory-6steps/` — representative layer/head attention heatmaps and trajectory contact sheets. |
| - `visualizations/density/trajectory-6steps/` — layer × step density heatmaps and layer-profile plots. |
| - `artifacts/videos/` — the `demo.mp4` output corresponding to the six-step run. |
| - `metadata/` — experiment table (`experiments.csv`), source checkout commit (`source-commit.txt`), complete file manifest with per-part checksums (`manifest.json`), and the raw archive checksum (`artifacts/archives/*.sha256`). |
|
|
| ## Reconstructing the archive |
|
|
| ```bash |
| cat lingbot-va-attn-trajectory-6steps.tar.gz.part* > lingbot-va-attn-trajectory-6steps.tar.gz |
| sha256sum lingbot-va-attn-trajectory-6steps.tar.gz |
| # expect: 672366873edd322eed50879226487042f38cbd3afbe7186ba584ef565b806c39 |
| tar -tzf lingbot-va-attn-trajectory-6steps.tar.gz | wc -l |
| # expect: 4329 (4,320 CSV + 1 summary.json + 8 directory entries) |
| ``` |
|
|
| ## Matrix semantics |
|
|
| Each raw `lXXhYY.csv` is a dense attention matrix; empty CSV cells encode masked/future columns and are preserved. The original `summary.json` inside the archive is authoritative for token/chunk boundaries and shifted scheduler timesteps (use `summary.json`, not the step index, for actual timesteps). The six-step trajectory comes from the `attn-exp-vv-10-6` run (`robotwin_i2av`, 10 chunks, 25 denoising updates) and is the LingBot-VA 1.0-style velocity FlowMatch execution path currently wired in this repository, not the VA2 distilled student. |
|
|
| This trimmed release intentionally omits the earlier full-history, smoke, and three-step captures and the duplicate `attn-exp.zip` archive. |
|
|
