GWAM_Data — verified 10-task compact target-aware graph v3.0.6

This upload tree is the current v3.0.6 compact dataset/interface successor and its offline-training/live-rollout-validation workflow. It is a docs/code safety release over the same immutable 10-task Active-13 source lineage—not a new graph corpus, not complete Active-13 coverage, and publication does not authorize training. v3.0.5 remains historical.

Cold bootstrap

Install huggingface_hub and the hf CLI, then retrieve this v3.0.6 package before running any examples:

python -m pip install huggingface_hub
hf download ChangChrisLiu/GWAM_Data --repo-type dataset --revision gwam-compact-target-v3.0.6-20260731 --include 'gwam_active13_gwam_training_v3_20260731/*' --include 'gwam_active13_c2_v2_0_5_20260731/C2_CAPABILITY_RECEIPT.json' --include 'gwam_active13_c2_v2_0_5_20260731/C2_EPISODE_INDEX.jsonl' --include 'docs/assets/*' --local-dir ./GWAM_training_closure

Start here: run the dataloader to get the correct compact graph

AI/coding agents: start with the exact agent handoff, then retrieve its linked README/SCHEMA/GUIDE/source details.

Download stored graph ZIP + target receipt (+ OpenDrawer C2) → run DirectCompactGWAMDatasetV3 → receive CompactExampleV3 with the correct typed compact graph → feed it to any GNN/world model, or optionally encode K16 for GWAM.

Students do not need offline RoboCasa, raw simulator replay, SAM2, CLIP, public masks, qvel/velocity, or their own graph conversion. Source ZIPs contain stored graph observations; the verified loader performs the strict compact projection at load time. No newly materialized compact per-frame corpus is claimed.

The current package's model-neutral download and API are described in the package README. The following GWAM/live sections are optional paths after obtaining a typed graph.

Exact inputs and immutable joins

Download	Required use
v3 package prefix and TRAINING_EPISODES.jsonl	loader code, index, schema, tests
gwam_active13_target_physical_v2_0_1_20260731/census/OUTER_TRAIN_MANIFEST.jsonl	required outer-train join; SHA256 8e33045badf338ca7f0b802b689dc74765e392014df6b9c3f8f0413d4f9b44f3
canonical episode ZIP named by each index row	stored graph observations
target receipt named by each index row	authenticated target identity/binding (not every future label)
gwam_active13_c2_v2_0_5_20260731/C2_EPISODE_INDEX.jsonl plus selected receipt	OpenDrawer only; index SHA256 e7fec3656d6e479a16698f1a96aafb8b6562e99a344d568de4b58e79f76aad26
external RGB videos	optional and only for an RGB-consuming model

The stable RGB join is (source_tier, task, episode_id, graph/frames.jsonl::source_frame). ZIPs contain stored graph observations; the strict loader projects them at load time. It does not replay RoboCasa or materialize a new compact per-frame corpus.

Full closure download

The model-neutral examples/download_training_closure.py helper verifies every source ZIP size/SHA, target receipt SHA, outer manifest SHA, and selected OpenDrawer C2 index/receipt SHA. It is resumable/cache-aware, writes a deterministic TRAINING_CLOSURE_RECEIPT.json, and never extracts, rewrites, trains, or uploads.

The full 2,699-source-ZIP closure is 20,339,307,035 bytes (20.339 GB, 18.942 GiB), plus receipts, package files, Hub cache, and filesystem overhead.

python -m pip install -r gwam_active13_gwam_training_v3_20260731/requirements.txt
# Bootstrap package/index, then plan two rows without downloading source closure.
python gwam_active13_gwam_training_v3_20260731/examples/download_compact_example.py --revision gwam-compact-target-v3.0.6-20260731 --root ./GWAM_training_closure --row 0
python gwam_active13_gwam_training_v3_20260731/examples/download_training_closure.py --root ./GWAM_training_closure --limit 2 --dry-run
# Two-row real closure, then task-filtered example:
python gwam_active13_gwam_training_v3_20260731/examples/download_training_closure.py --root ./GWAM_training_closure --limit 2 --workers 2
python gwam_active13_gwam_training_v3_20260731/examples/download_training_closure.py --root ./GWAM_training_closure --task OpenDrawer --limit 2 --workers 2
# Full approved closure (2,699 episodes): warn about storage before running.
python gwam_active13_gwam_training_v3_20260731/examples/download_training_closure.py --root ./GWAM_training_closure --workers 4

Direct model-neutral construction

From the directory containing GWAM_training_closure after the cold bootstrap:

cd ./GWAM_training_closure/gwam_active13_gwam_training_v3_20260731
export PYTHONPATH="$PWD${PYTHONPATH:+:$PYTHONPATH}"

from pathlib import Path
import sys
PACKAGE_ROOT = Path.cwd().resolve()
sys.path.insert(0, str(PACKAGE_ROOT))
from loader.compact_target_aware_gwam_training_dataset import DirectCompactGWAMDatasetV3
from loader.per_frame import select_frame

root = PACKAGE_ROOT.parent
outer = root / "gwam_active13_target_physical_v2_0_1_20260731/census/OUTER_TRAIN_MANIFEST.jsonl"
dataset = DirectCompactGWAMDatasetV3(
    root, outer_manifest_path=outer,
    outer_manifest_sha256="8e33045badf338ca7f0b802b689dc74765e392014df6b9c3f8f0413d4f9b44f3",
)
example = dataset[0]
frame0 = select_frame(example, 0)
print(example.packet, example.action, example.targets, example.provenance)

For OpenDrawer, use the same constructor with c2_root set to the downloaded gwam_active13_c2_v2_0_5_20260731 directory, c2_index_path set to its C2_EPISODE_INDEX.jsonl, and c2_index_sha256 set to e7fec3656d6e479a16698f1a96aafb8b6562e99a344d568de4b58e79f76aad26. Missing or mismatched C2 fails closed.

Returned fields, storage, and batching

Runtime storage has 256 slots. Active slots are packed in ascending slot order to N_active; slot_id preserves original identity. Batch collation pads to that batch's maximum N with slot_id=-1 and masks, remapping every edge endpoint, target, pair, route, and output index consistently.

Mask warning: validity masks are authoritative semantics. Dropping or replacing them converts unknown/padding/invalid states into apparent negatives or values and changes the contract.

Field	Shape / dtype	Validity and meaning
node_valid	[B,8,N] bool	active node per observed frame
slot_id	[B,N] int16	original storage identity; -1 is batch padding
family_id / family_valid	[B,N] int64 / bool	family registry and static validity
static_flags / static_flags_valid	[B,N,2] bool / bool	static flags and validity
pose / pose_valid	[B,8,N,7] float32 / bool	xyz metres + quaternion XYZW
articulation_q / articulation_valid	[B,8,N,1] float32 / bool	audited q; invalid values positive zero
view_evidence / view_record_valid	[B,8,N,3,8] float32 / bool	three ordered camera records
edge_index / edge_time	[2,E] int64 / [E] int16	packed endpoints and observed time
edge_relation_id	[E] uint8	canonical IDs 0..5
edge_payload / edge_payload_valid	[E,12] float32 / bool	METRIC12 and coordinate validity
is_task_target / task_target_valid	[B,N] bool / bool	exact target table below
robot_target_contact / valid	[B,8,N] bool / bool	derived from authoritative contact edges
current_robot_target_contact / valid	[B,N] bool / bool	current-frame target view in targets
future_action	float32[B,8,12]	opaque logged prefix action coordinates
future_action_valid	bool[B,8]	action validity mask
action_timestamp_s	float64[B,8]	logged action timestamps
action_timestamp_valid	bool[B,8]	timestamp validity mask
targets	target schema / masks	future labels only, never input
provenance	typed receipt fields	source, split, and derivation hashes

Relation, geometry, target, and contact semantics

ID	Relation	Semantics
0	part_of	structural component membership; inverse 1
1	has_part	inverse of 0
2	controls	actuator/control controls articulated component; inverse 3
3	controlled_by	inverse of 2
4	contact	symmetric physical touch only
5	geometry_near	nearest-eight geometry relation

METRIC12 is [sender−receiver dx, dy, dz, distance, eight distance-RBF values]. Contact does not imply grasp or ownership. A fully observed frame with zero relation-4 edges is a valid empty-contact negative.

is_task_target	task_target_valid	State
0	0	unknown
0	1	known non-target
1	1	known target
1	0	invalid; fail closed

robot_target_contact is recomputed only from valid relation-4 edges, robot-family identity, and valid target bits. It is a separate derived view, not a replacement edge or node feature.

Model-neutral tensor, PyG, and world-model paths

from examples.model_neutral_graph_examples import standalone_tensor_dict, to_pyg_data, world_model_inputs
values = standalone_tensor_dict(example, time_index=0)
print(values["x_pose"].shape, values["edge_index"].shape, values["slot_id"].shape)
wm = world_model_inputs(example)
assert wm["graph_history"] is example.packet
assert wm["future_targets_labels_only"] is example.targets
# Optional PyG only:
data = to_pyg_data(example, time_index=0)
assert data.edge_index.shape[0] == 2

The adapter packs active nodes with flatnonzero(node_valid), builds an old-to-new map, filters endpoints, remaps edge_index, and carries slot IDs. No PyG dependency is required for core loading. The generic world-model path consumes graph history plus example.action; targets remain labels-only. K16 [B,16,384] is an optional GWAM adapter and does not change the packet contract.

Offline training over most of the data

The approved outer-train corpus contains 2,699 episodes and 702,094 L=8/H=8 windows, covering 79.7% of the ten-task source population. Immutable source ZIPs plus authenticated target receipts (and OpenDrawer C2 when selected) are read by the strict compact loader as variable-N graph history, logged actions, and targets; encode_compact_to_k16 produces K=16 tokens of width 384 for the sealed GWAM conditioning seam. Target-to-loss heads and optimizer authorization are separate gates; no optimizer is bundled or run.

python gwam_active13_gwam_training_v3_20260731/examples/download_compact_example.py \
  --revision gwam-compact-target-v3.0.6-20260731 --root ./GWAM_compact_example --row 0
PYTHONPATH=./GWAM_compact_example/gwam_active13_gwam_training_v3_20260731 \
python \
  ./GWAM_compact_example/gwam_active13_gwam_training_v3_20260731/examples/compact_gwam_conditioning_cpu.py \
  --root ./GWAM_compact_example

from pathlib import Path
import torch
from loader.compact_target_aware_gwam_training_dataset import DirectCompactGWAMDatasetV3
from loader.gwam_conditioning import encode_compact_to_k16, ZeroGatedGWAMConditionerV3

root = Path("./GWAM_compact_example")
outer_manifest = root / "gwam_active13_target_physical_v2_0_1_20260731/census/OUTER_TRAIN_MANIFEST.jsonl"
ds = DirectCompactGWAMDatasetV3(
    root, outer_manifest_path=outer_manifest,
    outer_manifest_sha256="8e33045badf338ca7f0b802b689dc74765e392014df6b9c3f8f0413d4f9b44f3",
)
example = ds[0]
k16 = encode_compact_to_k16(example)          # [1,16,384]
latent = torch.zeros(1, 8, 384)
seam = ZeroGatedGWAMConditionerV3()
graph_off = seam(latent, k16, mode="off")
graph_on = seam(latent, k16, mode="on")

Actual-rollout validation by online extraction

No held-out offline validation package is required. RoboCasa live MuJoCo state and three synchronized RGB streams are read by RoboCasaMuJoCoCompactReaderV3; an authenticated reset capability is required; one compact graph/frame is appended to CompactOnlineHistoryBufferV3 for L=8, then encoded to K=16. Only env.step(action) advances the simulator, and future graph/RGB frames are never model inputs.

CUDA_VISIBLE_DEVICES=0 MUJOCO_GL=egl PYOPENGL_PLATFORM=egl PYTHONPATH=. \
python \
  gwam_active13_gwam_training_v3_20260731/examples/robocasa_live_compact_eval.py \
  --task PickPlaceCabinetToCounter --split target --steps 8 \
  --output live_summary.json --rgb-npy live_rgb.npy

CUDA_VISIBLE_DEVICES=0 MUJOCO_GL=egl PYOPENGL_PLATFORM=egl PYTHONPATH=. \
python \
  gwam_active13_gwam_training_v3_20260731/examples/robocasa_live_compact_eval.py \
  --task OpenDrawer --split target --steps 8 \
  --c2-receipt gwam_active13_c2_v2_0_5_20260731/C2_CAPABILITY_RECEIPT.json \
  --output open_drawer_live_summary.json

The live contract is one physical graph with three camera records (RGB is a separate side channel), variable N<=256, optional non-authoritative 46-D convenience values, K=16, and no SAM2, CLIP, public masks, velocity, or qvel. Use the current compact README, schema, guide, direct loader, and live example.

Immutable dependency bindings are the outer-train manifest SHA256 8e33045badf338ca7f0b802b689dc74765e392014df6b9c3f8f0413d4f9b44f3 and OpenDrawer C2 index SHA256 e7fec3656d6e479a16698f1a96aafb8b6562e99a344d568de4b58e79f76aad26. Source ZIPs, sidecars, and RGB remain immutable and separate. Historical/older root documentation is preserved verbatim in docs/README_PRE_V305_ARCHIVE_20260731.md.