| ---
|
| license: other
|
| tags:
|
| - counter-strike
|
| - pov
|
| - world-model
|
| ---
|
|
|
| # CS2 single-POV HLTV artifacts (heads, latents, labels)
|
|
|
| The irreplaceable ~183 MB extracted from a 163 GB local render workspace, so the
|
| bulk could be deleted. Companion to
|
| [cs2-hwm-scale300](https://huggingface.co/cbctr/cs2-hwm-scale300) and
|
| [cs2-10k-vjepa2-latents-300](https://huggingface.co/datasets/cbctr/cs2-10k-vjepa2-latents-300).
|
|
|
| These come from rendered GOTV POV of 19 pro matches. Unlike CS2-10k, **these
|
| renders contain HUD and radar pixels**, which is what makes the HUD digit reader
|
| and the economy/value heads possible at all - CS2-10k cannot support that line
|
| of work.
|
|
|
| ## Contents
|
|
|
| | path | what |
|
| |---|---|
|
| | `planner_bundle/*.pt` | 10 trained heads: HUD digit + token readers, value, option value, destination, enemy belief (audio-only / video-only / video+audio), utility threat |
|
| | `pov_vjepa_cache/` | frozen V-JEPA2 POV latents + `repair_causal_vjepa_ckpt.pt` |
|
| | `pov_observations.parquet` | per-window observation labels and round-win targets |
|
| | `pov_multimodal_manifest.parquet` | window -> video segment/frame mapping |
|
| | `option_value_cache.parquet` | **the train/holdout split** the caches were built against |
|
| | `pov_economy_labels.parquet`, `destination_labels.parquet` | economy + destination labels |
|
| | `pov_hud_digit_tokens.parquet`, `pov_hud_tokens.parquet` | HUD reader outputs |
|
| | `destination_zones.json` | 8-zone map partition (centroids, names, roles) |
|
| | `*_report.json` | measured results for each head |
|
|
|
| ## What is NOT here
|
|
|
| The 143 GB of rendered POV mp4 segments, the HUD crop/digit image caches, and
|
| the source `.dem` files. Those are regenerable from HLTV demos via the CSDM
|
| render pipeline (~52 h), assuming the demos are still hosted.
|
|
|
