StreamVLN SatNav Scratch 1ep f32h8s4

This repository contains a Hugging Face upload-ready copy of the StreamVLN SatNav scratch baseline streamvln-baseline-scratch-1ep-f32h8s4-lr2e-5. The model predicts navigation actions from visual observations and a language instruction using the StreamVLN/Video-LLM action-generation code path.

Model Details

Model type: StreamVLN-style vision-language-navigation policy.
Training mode: scratch SatNav fine-tuning from lmms-lab/LLaVA-Video-7B-Qwen2.
Vision tower source: google/siglip-so400m-patch14-384.
Precision: bfloat16 weights stored as sharded safetensors.
Window setting: f32h8s4 means 32 frames, 8 history steps, and 4 future steps.
Local archive name: streamvln-baseline-scratch-1ep-f32h8s4-lr2e-5.

The vision tower was included during training (mm_tunable_parts=mm_vision_tower,mm_mlp_adapter,mm_language_model), and its fine-tuned parameters are saved inside the main model-*.safetensors shards under model.vision_tower.*. The google/siglip-so400m-patch14-384 value in config.json identifies the SigLIP base architecture/source used to construct the tower before checkpoint weights are loaded.

Repository Contents

This release directory contains only inference/evaluation artifacts:

sharded safetensors model weights
model.safetensors.index.json
model and generation configs
tokenizer files
model card, license, and notice files

DeepSpeed optimizer state, scheduler state, RNG state, trainer state, training logs, TensorBoard/SwanLab outputs, and duplicated checkpoint-* directories are intentionally omitted.

Training Data and Procedure

The model was fine-tuned for 1 epoch on SatNav trajectory data using the StreamVLN SatNav training pipeline. The default training setup used:

SatNav dataset version: SatNav-v0.1
frames/history/future: 32 / 8 / 4
optimizer learning rate: 2e-5
full-parameter fine-tuning
bfloat16 training

The model card does not redistribute SatNav episodes, images, simulator scenes, or training logs.

License

This release is licensed as cc-by-nc-sa-4.0. This conservative choice follows the upstream StreamVLN checkpoint release terms for redistribution of StreamVLN-family model weights.

Third-party components remain under their own terms, including lmms-lab/LLaVA-Video-7B-Qwen2, google/siglip-so400m-patch14-384, the StreamVLN codebase, SatNav data/assets, simulator assets, and external code dependencies. SatNav episodes, images, simulator scenes, optimizer state, and training logs are not redistributed in this repository.

Before making the repository public, confirm that the SatNav dataset and any project-specific release requirements allow public redistribution of the trained weights.