spamod-ckpts-1

SpatialDrive Checkpoints (part 1 of 2)

Training checkpoints from the SpatialDrive project: a BEV-augmented LLaVA-OneVision (SigLIP-so400m-patch14-384 vision tower + Qwen2-7B-Instruct LM, model_type: llava_qwen with added BEV compressor/projector modules) fine-tuned for spatial reasoning on multi-camera autonomous-driving data (nuScenes).

• Part 1 (this repo): runs Sep 08 – Oct 31 (stage 1/2/3/6 chain)
• Part 2: ktian6/spamod-ckpts-2 (Nov runs, BEV/interleaved/LoRA + final merged models)
• Code: https://github.com/kxint9625/spamod (training yamls in LLaVA-NeXT/scripts/train/, entry scripts stage*.sh)
• BEV feature extractor checkpoints: ktian6/spamod-bevdet

Each folder is a complete standalone HF-format model (final weights of that run; intermediate checkpoint-XXXX trainer states were not uploaded). config.json:_name_or_path still records the original local path of the run each model was initialized from.

Run lineage (this repo)

Folder prefix llava-onevision-google_siglip-so400m-patch14-384-Qwen_Qwen2-7B-Instruct-ov is abbreviated to ov below. Runs marked † were initialized from a run that was already deleted before archiving (not recoverable).