Document unified feature-only streaming protocol

README.md

@@ -1,12 +1,57 @@
 ---
-license: other
-pretty_name: csgo-training-cache
+language:
+- en
+- zh
+tags:
+- video
+- world-model
+- csgo
+- training-cache
+- streaming
 ---
+# CSGO training cache
 
-# csgo-training-cache
+Precomputed feature-only training samples, continuously converted from four source datasets:
 
-CSGO training cache: frozen Wan2.1-VAE features (`latent`, `i2v_y`, `dense_features`) plus `state`,
-`interaction`, `camera_poses`, `camera_intrinsics`, `text_emb`, converted from `peih/csgo-4090-v4-20260921`. Raw Dense7 is not included.
+- `peih/csgo-3090-v4-overflow-20260923`
+- `peih/csgo-4090-v4-20260921`
+- `peih/csgo-4090-v4-overflow-20260923`
+- `peih/csgo-dual3090-v4-backup`
 
-One TAR per source archive (one match) under `shards/<pod>/`; a shard is valid only once
-`receipts/<pod>/<archive_sha>.json` exists. Consume through the versioned manifest in `peih/csgo-training-cache`, never by listing this repo.
+This dataset is still growing. `manifest.json` is the current verified snapshot; `manifests/vNNNNNN.json` contains immutable historical snapshots. Read only listed shards at their pinned `revision`. New uploads become available after their verification receipt and a manifest update. Do not train from unlisted TAR files.
+
+## Format
+
+Schema: `csgo_training_cache_features_v1`. Each TAR contains ordered pairs `<sample_id>.pt` and `<sample_id>.json`. A shard usually holds 60 samples and is approximately 1.49 GB; use the exact size in the manifest.
+
+| Tensor | Shape per sample | dtype |
+|---|---|---|
+| latent | [1,16,21,60,104] | float32 |
+| i2v_y | [1,20,21,60,104] | float32 |
+| dense_features | [1,21,32,30,52] | float32 |
+| state | [1,21,3,60,104] | float32 |
+| interaction | [1,21,134] | float32 |
+| camera_poses | [1,21,4,4] | float32 |
+| camera_intrinsics | [1,21,4] | float32 |
+| text_emb | [1,L,4096] | bfloat16 |
+
+There is **no raw seven-channel `dense` tensor**. The Dense VAE features are already encoded. No VAE/T5 encoding is required on the training consumer. Model revisions, weight hashes and preprocessing contract are recorded in the manifest.
+
+Use `torch.load(..., weights_only=True)`. Verify each PT's bytes and SHA256 before deserialization, and verify the complete TAR SHA256 at EOF. A partially consumed shard has not yet passed the consumer's full-TAR check.
+
+## Streaming consumer contract
+
+1. Read `manifest.json` at a fixed dataset commit.
+2. Respect train/val splits and `exclude_sample_ids`. The split is stable by match, with an expected validation fraction of 2% (not an exact per-snapshot ratio).
+3. Open listed TARs over streaming HTTP at each pinned shard revision. Avoid downloading the entire dataset.
+4. Interleave shards; assign disjoint shards to different ranks. Sample IDs must not repeat across ranks in the same epoch.
+5. Keep the manifest fixed throughout an epoch. Refresh only at the next epoch boundary to incorporate newly published samples.
+6. Checkpoint the manifest, deterministic order/cursor and training state. A restart may reread a TAR prefix from the network but must not train already consumed samples again.
+
+This layout uses a manifest-aware custom TAR reader. Ordinary `datasets.load_dataset(streaming=True)` is not claimed to implement the exclusion, version refresh and exact-resume protocol automatically.
+
+## Publication
+
+Two conversion machines write separate `shards/<worker>/...` and `receipts/<worker>/...` paths in this repo. One manifest publisher serializes publication with a parent-commit check. Existing verified shards from the four older cache repos are being copied byte-for-byte; this is not re-encoding.
+
+The full dataset is not yet complete. Manifest totals are the authority for currently published samples. Source discovery and conversion continue while consumers train on a fixed snapshot.