SimpleMemVLN R2R fullcontext — option B
Not yet evaluated in Habitat. Training loss is not a navigation success metric. The observation-before-action alignment is declared but not independently verified from collector source or Habitat replay.
Recipe
Independently initialized from Qwen/Qwen3.5-4B revision
851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, not from the other variant.
Code commit: d50174daf1fda692849cfaa3e69286a504339fb2.
Serializer: vln_append_only_chat_v1. Native LLM text-action head (option B),
gold action history, per-action mean token CE including assistant terminator.
Frozen vision/merger; trainable text backbone and LLM head. BF16, ZeRO-2,
nonreentrant full-episode gradient checkpointing, no truncation or TBPTT.
Four H100 GPUs, one episode/rank, GAS 2: global batch 8 episodes.
R2R train: 10,819 unique episodes, 631,244 actions; distributed repetition
adds one episode per epoch. Three epochs, 4,059 updates, 122 warmup updates.
Peak LR 5e-6, cosine decay to 10% of peak, weight decay 0.01, seed 429.
Memory mode: fullcontext. Window8 keeps prefix plus eight complete steps
for full-attention KV; persistent GDN state and gradients are not reset.
Snapshots
epoch-1, epoch-2, epoch-3 correspond to updates 1353, 2706, 4059.
Epochs 1–2 are mid-schedule snapshots of one 3-epoch cosine run, not
independently trained one-/two-epoch models. All optimizer states remain local.
Loading and integrity
These weights use the SimpleMemVLN navigation wrapper, not an unmodified
AutoModel state dictionary. Download an epoch directory and load it with
qwen_vl.train.vln_runtime.load_checkpoint(path, model_path) using the pinned
base snapshot and the recorded environment. Navigation metadata contains the
complete serializer and memory contract.
SHA256SUMS.json records model/configuration/metric hashes. The uploader
downloads every uploaded file at the immutable commit revision and verifies
SHA256 before marking publication complete locally. Dataset images, credentials
and optimizer states are not published.