Model Summary
This repository contains a SwiftVLN 3B checkpoint for SatNav aerial visual-language navigation evaluation. The checkpoint is based on the Qwen2.5-VL 3B architecture and is packaged as Hugging Face inference/evaluation artifacts for the SwiftVLN codebase.
This model is part of the SwiftVLN SatNav 3B Model Zoo, a set of eleven 1-epoch checkpoints covering PF, MAP, GTC/SGTC, overlap ablations, initialization/randomization settings, and PoseFiLM/no-embedding configurations.
Variant
Table with columns: Field, Value| Field | Value |
|---|
| Short eval name | swiftvln-satnav-3b-1ep-f32s4-overlap0-map-g1000-l400-r448-d20-s2-noembed |
| Source checkpoint | swiftvln-satnav-3b-1ep-f32s4-overlap0-map-g1000-l400-r448-d20-s2-noembed-bs64-lr2e-5-092810 |
| Base architecture | Qwen2.5-VL 3B |
| Training task | SatNav aerial visual-language navigation |
| Training schedule | 1 epoch, f32s4, batch size 64, learning rate 2e-5 |
| Variant note | MAP variant with g=1000, l=400, r=448, d=20, s=2, and no additional embedding slot. |
| Weight format | safetensors |
| Referenced weight shards | 2 |
The abbreviations in the checkpoint name, such as PF, MAP, GTC, SGTC, PoseFiLM, overlap, pool, and noembed, refer to SwiftVLN experiment settings. Use the short eval name above when running the repository's evaluation scripts.
Intended Use
This checkpoint is intended for research evaluation of SwiftVLN on SatNav navigation episodes. It is most directly used with the SwiftVLN evaluation entrypoint:
MODEL_PATH=Eku127/swiftvln-satnav-3b-1ep-f32s4-overlap0-map-g1000-l400-r448-d20-s2-noembed bash src/swiftvln/scripts/eval/eval_by_name.sh swiftvln-satnav-3b-1ep-f32s4-overlap0-map-g1000-l400-r448-d20-s2-noembed
The model is not intended to be used as a general-purpose chat, VQA, robotics, or real-world autonomous navigation model. It expects the prompt format, observation construction, and SatNav evaluation environment used by SwiftVLN.
Files
This upload-ready repository keeps only the files needed for Hugging Face loading and SwiftVLN evaluation:
config.json, generation config, tokenizer files, processor/preprocessor configsmodel.safetensors.index.json- referenced
model-*.safetensors weight shards
.gitattributes and this model card
Training-only artifacts such as optimizer state, scheduler state, trainer state, RNG state, logs, TensorBoard runs, checkpoint wrapper directories, and plots are intentionally excluded.
Evaluation
For by-name evaluation inside the SwiftVLN repository, use:
MODEL_PATH=Eku127/swiftvln-satnav-3b-1ep-f32s4-overlap0-map-g1000-l400-r448-d20-s2-noembed bash src/swiftvln/scripts/eval/eval_by_name.sh swiftvln-satnav-3b-1ep-f32s4-overlap0-map-g1000-l400-r448-d20-s2-noembed
The SwiftVLN evaluation script resolves the model from MODEL_PATH and reads the SatNav eval configuration from the repository. Typical evaluation uses the SatNav val_seen and val_unseen splits configured in the local SwiftVLN setup.
Limitations
- This checkpoint is specialized for SatNav aerial navigation and has not been validated for general image-text tasks.
- Correct use depends on the SwiftVLN prompt format, preprocessing, action representation, and SatNav episode data.
- The repository does not include the SatNav dataset or evaluation environment.
- The model card does not claim safety or reliability for real-world autonomous operation.
- No repository-specific license is declared here; downstream users should check the applicable licenses and terms for the base model, data, and code they use with this checkpoint.
Citation
If you use this checkpoint, cite the SwiftVLN/SatNav project or associated paper/report from the code release that provided the model zoo.