YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
M5_L — learned GDN memory rankers on M5 (SimpleMemVLN, Phase B)
Each top-level folder is one completed navigation model (L0/ … L3/). The arms change only how M5's
episodic memory ranks old visual K/V rows for its fixed R4/B1200 budget: a small frozen predictor per
full-attention layer, fitted offline to exact leave-one-row-out attention-output utilities, scores each row
from birth-time Gated DeltaNet features. Everything else matches the M5_E1 reference
(anhdao69/Streaming_VGGT, folder M5_E1, the fixed-score S0 arm) and the fixed-score sweep
(anhdao69/M5_S).
| Arm | Model | Features |
|---|---|---|
| L0 | Linear(96 → 1) | F0_v1 of the GDN layer before each attention layer |
| L1 | MLP 96 → 64 → 1 | F0_v1, previous GDN layer |
| L2 | MLP 288 → 64 → 1 | F0_v1, previous three GDN layers |
| L3 | MLP 576 → 64 → 1 | F1_v2, previous three GDN layers |
Each folder holds the serving files (the frozen ranker is embedded in navigation.json), README.md
(arm definition, offline ranking quality, training summary, loading), provenance/ (identity,
qualification, ranker ledger, label summary, drift and validation evidence) and SHA256SUMS.json.
Folders are added one at a time as training finishes; earlier folders are never modified.
Code: SimpleMemVLN branch m5_gdn_l_sweep. Training loss and offline ranking quality are reported per arm;
navigation success has not been evaluated in these folders.