Remove duplicated release-prefix root mirror files
Browse files
gwam_v12_sparse_v2/docs/QA_AND_LEAKAGE.md
DELETED
|
@@ -1,66 +0,0 @@
|
|
| 1 |
-
# GWAM_Data v1.2 sparse-v2 QA and leakage guards
|
| 2 |
-
|
| 3 |
-
## Required shape checks
|
| 4 |
-
|
| 5 |
-
```text
|
| 6 |
-
node_state: float32[T, 256, 32]
|
| 7 |
-
node_state_schema: str[32]
|
| 8 |
-
view_evidence: must agree with visible mask pair set
|
| 9 |
-
dynamic_edges: sparse JSONL, contact + visibility_change only
|
| 10 |
-
Phase 2: sparse visible-only features; no dense visual tensor required
|
| 11 |
-
```
|
| 12 |
-
|
| 13 |
-
Reject stale v1/v1.1 expectations such as 317/323-D node features, dense `rel_state`, dense `visual_views`, or `distractor` family index 1.
|
| 14 |
-
|
| 15 |
-
## Phase 1 checks
|
| 16 |
-
|
| 17 |
-
- `_PHASE1_DONE.json` exists for every packaged episode.
|
| 18 |
-
- `graph_static.json` schema is `GWAM_Data_v1.2_phase1_sparse_v2`.
|
| 19 |
-
- All packaged episodes share one identical `rule_versions` blob.
|
| 20 |
-
- `node_state` has no NaN/Inf.
|
| 21 |
-
- Padded rows are all-zero; active rows are selected via the active channel.
|
| 22 |
-
- `view_evidence` visible pairs exactly match RLE mask visible pairs.
|
| 23 |
-
- Dynamic-edge indices are valid and edge vocab is only `contact` / `visibility_change`.
|
| 24 |
-
- Actions are present and non-null in `frames.jsonl`.
|
| 25 |
-
|
| 26 |
-
## Phase 2 checks
|
| 27 |
-
|
| 28 |
-
For every packaged episode with Phase 2:
|
| 29 |
-
|
| 30 |
-
- `audit/phase2_summary.json` exists.
|
| 31 |
-
- No invisible features are injected.
|
| 32 |
-
- Valid and invalid sparse visual-feature pairs partition the Phase 1 visible pair set.
|
| 33 |
-
- Valid feature vectors are fp16 and unit-normalized.
|
| 34 |
-
- No duplicate feature pairs.
|
| 35 |
-
|
| 36 |
-
## Leakage guards
|
| 37 |
-
|
| 38 |
-
Model-input tensors must not contain task-goal labels or target-relevance bits.
|
| 39 |
-
|
| 40 |
-
Default model inputs may use:
|
| 41 |
-
|
| 42 |
-
```text
|
| 43 |
-
node_state active rows
|
| 44 |
-
sparse dynamic_edges
|
| 45 |
-
Phase 2 visible-only features and their pair indices
|
| 46 |
-
action/robot streams when explicitly included
|
| 47 |
-
```
|
| 48 |
-
|
| 49 |
-
Treat these as metadata/supervision-only unless explicitly allowed:
|
| 50 |
-
|
| 51 |
-
```text
|
| 52 |
-
instruction
|
| 53 |
-
fixture_refs
|
| 54 |
-
object_roles
|
| 55 |
-
goal metadata
|
| 56 |
-
node_id strings
|
| 57 |
-
source strings
|
| 58 |
-
frames.jsonl timestamp/index/next.done
|
| 59 |
-
```
|
| 60 |
-
|
| 61 |
-
Node ordering can correlate with object config order; do not use raw node slot position as semantic signal unless studying that bias explicitly.
|
| 62 |
-
|
| 63 |
-
## Known documented exceptions
|
| 64 |
-
|
| 65 |
-
- `target/OpenDrawer/episode_000459` is excluded due source env-init failure.
|
| 66 |
-
- Six target tasks are state-blind in `node_state`: CloseFridge, OpenStandMixerHead, SlideDishwasherRack, TurnOffStove, TurnOnElectricKettle, TurnOnMicrowave. They need a future `fixture_state.npz` sidecar or narrower dataset-card claims.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|