# VLWM Training Data Release Status

Fetched: 2026-07-15

## Canonical Links

- Paper: https://arxiv.org/abs/2509.02722v2
- Official Hugging Face dataset placeholder: https://huggingface.co/datasets/delong-chen/VLWM
- Official Hugging Face model page: https://huggingface.co/delong-chen/VLWM
- Author announcement: https://x.com/Delong0_0/status/1963883798841483688

## Public Access Status

The Hugging Face dataset repository exists but is empty. Its only files are `.gitattributes` and a 56-byte `README.md` containing the `fair-noncommercial-research-license` frontmatter field. The Hugging Face API reports `usedStorage: 0`; there are no data files, configs, splits, rows, schema, or Dataset Viewer records.

The model repository exists and is manually gated. Its public file listing contains only `.gitattributes` and `README.md`; the public landing page does not expose model weights. The paper source contains a commented metadata URL for `https://github.com/facebookresearch/VLWM`, but that repository returned 404 during this ingest and is not treated as a released artifact.

## What The Paper Says The Training Targets Contain

Each example predicts:

1. a concise goal description;
2. a goal interpretation describing initial and expected final world state;
3. an interleaved sequence of action descriptions and textual world-state changes.

The video-derived targets are generated from hierarchical Trees of Captions using Llama-4 Maverick and two rounds of Self-Refine. A text-only branch converts NaturalReasoning chains of thought into the same action/state trajectory format.

## Paper-Reported Sources

The implementation table lists NaturalReasoning, HowTo100M, COIN, CrossTask, YouCook2, EgoExo4D, and EPIC-KITCHENS-100. The introduction also names Ego4D, but the implementation section says Ego4D was excluded to avoid benchmark overlap and the table omits it.

The table's component rows sum to 2,237.3k trajectories and 10,905.8k steps, while its Overall row reports 2,179.6k trajectories and 10,604.3k steps. The public empty repository cannot resolve this arithmetic discrepancy.

## Relationship To Action100M

Action100M is a later, public annotation release from an overlapping author team. Its paper explicitly says that its pipeline extends and improves the data-generation procedure introduced by VLWM. It is **not** the exact VLWM training corpus:

- Action100M is built from HowTo100M only, whereas VLWM reports six video sources plus NaturalReasoning.
- Action100M provides hierarchical segment captions and action labels; it does not expose VLWM's goal, goal-interpretation, and explicit textual post-action state-change target schema.
- Only a 10% Action100M preview is currently published on Hugging Face.

See `datasets/action100m-2026/` for the released dataset metadata.