SimpleMemVLN FullContext Candidate Logits — No Action History
Not yet evaluated in Habitat. Training loss is not navigation success.
Completed one-epoch joint R2R + English RxR_15deg experiment, job 4652.
epoch-1/ contains update 3852, the end of the complete one-epoch cosine
schedule, not an intermediate snapshot of a two-epoch run.
Architecture and contract
Full-context visual history with persistent GDN state. Four-way candidate-logit readout uses selected pretrained, trainable Qwen LM-head rows for A/B/C/D, mapping to MOVE_FORWARD / TURN_LEFT / TURN_RIGHT / STOP. No autoregressive action generation or random classifier. Four-way cross entropy; class weighting none.
No action history: neither predicted nor teacher-forced action content is fed back. Class-independent assistant closure remains. Visual history is retained. Serializer: vln_candidate_logits_no_action_history_v1. This contract differs from both qwen_text and candidate-token-history checkpoints.
Recipe
30,815 episodes: 10,819 R2R and 19,996 English RxR_15deg; 3,128,624 actions.
Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
Frozen vision encoder, trainable language backbone and selected tied LM-head rows.
Global episode batch 8 = 4 H100 GPUs x 1 episode/rank x accumulation 2.
One epoch, 3852 updates, 116 warmup updates, backbone LR 5e-6, weight decay 0.01,
seed 429. Exact resolved recipe in provenance/recipe.yaml.
Source commit: 5d8edfd33c8213086464ad4243755a50ea09614f on streaming_logits.
Longest-trajectory preflight: 627 frames, peak reserved 76.8262 GiB.
Loading and integrity
Use the matching SimpleMemVLN source and its navigation wrapper loader:
from qwen_vl.train.vln_runtime import load_checkpoint
model, serializer = load_checkpoint("downloaded_repo/epoch-1", pinned_base_model_path)
These are navigation-wrapper weights, not a plain AutoModel state dictionary.
The pinned Qwen base model/processor is required. Weights, processor/tokenizer,
navigation metadata, training-state summary, provenance and epoch report are included.
Optimizer tensors, RNG state, dataset content and credentials are excluded.
SHA256SUMS.json records file hashes; publication is verified by downloading
every file from the immutable commit and checking SHA256.