SimpleMemVLN FullContext dual-lane, no action history — adapted
epoch-2/ is the completed two-epoch model at update 7704. The
first-epoch snapshot remains available under epoch-1/.
Initialized by adapting the completed trained joint no-action-history FullContext policy, with fresh step lanes. This snapshot counts adaptation epochs after that parent policy.
The adaptation schedule was extended from one to two total epochs at checkpoint 100. Weights, optimizer, scheduler and RNG state were preserved. The original 116-step warmup was retained; cosine decay now targets update 7704. The epoch-1 wall-clock report includes the interruption before recovery.
Architecture and training
Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
FullContext attention with parallel step lanes at layers 16, 20, 24 and 28,
using the step_end clock. Four-way candidate logits from pretrained trainable
LM-head rows; no appended action-history tokens.
Serializer: vln_candidate_logits_no_action_history_v1.
Frozen vision encoder and merger, trainable language backbone and step lanes.
Complete trajectories, gradient checkpointing, CPU activation offload for long
sequences, BF16 and DeepSpeed ZeRO-2.
Joint R2R + English-guide RxR_15deg: 30,815 episodes (10,819 + 19,996). Global batch 8 = four H100 GPUs × one episode/rank × accumulation 2. Backbone/head LR 5e-6; step-lane LR 1e-4; warmup 116 updates; cosine decay to 10% of peak, weight decay 0.01, seed 429. Epoch-1 action-weighted training loss: 0.2250577306. Epoch-2 action-weighted training loss: 0.0346941991. These are training metrics; this snapshot has not been evaluated in Habitat.
Loading and provenance
Use SimpleMemVLN source commit 3c8f580063d4fedb9ba3fdce98431f1e9a25e137 and the pinned runtime documented
in the streaming_logits_dual branch. Download the epoch directory and use:
from qwen_vl.train.vln_runtime import load_checkpoint
model, serializer = load_checkpoint("downloaded_repo/epoch-2", pinned_base_model_path)
The navigation wrapper and step-lane weights are included in pytorch_model.bin.
This requires the SimpleMemVLN loader and the pinned base model/processor;
an unmodified AutoModel loader does not implement this navigation contract.
navigation.json records the complete recipe and architecture.
Model, tokenizer/processor, epoch report and selected provenance are published.
Optimizer/RNG tensors, dataset contents and credentials remain local.
SHA256SUMS.json records file hashes. Every uploaded file is downloaded from
the immutable publication commit and checked against its local SHA-256.