ktian6/spamod-ckpts-2

SpatialDrive Checkpoints (part 2 of 2)

Training checkpoints from the SpatialDrive project (BEV-augmented LLaVA-OneVision; see part 1 for the project description and the earlier half of the lineage).

The final models of the project are in merged/:

• merged/..._1111_stage_lora2 — the last and most-refined full model (LoRA stage 2 merged, BEV-enabled)
• merged/..._1110_stage_lora1 — its predecessor (LoRA stage 1 merged)

Code: https://github.com/kxint9625/spamod · BEV extractor: ktian6/spamod-bevdet

Run lineage (this repo)

Prefix llava-onevision-google_siglip-so400m-patch14-384-Qwen_Qwen2-7B-Instruct-ov → ov. † = parent run deleted before archiving. Chain continues from part 1's ov_1031_stage_6_ad5.

Run	Initialized from	Steps	Epochs	Notes
ov_1101_stage_6_b12	1101_stage_6_b11 †	2811	3
ov_1106_stage_6_inter1	ov_1101_stage_6_b12	1770	2	interleaved
ov_1109_stage_6_inter2	ov_1106_stage_6_inter1	885	1	interleaved
ov_1109_stage_6_inter3	ov_1109_stage_6_inter2	250	1	interleaved
ov_1106_stage_6_inter6	ov_1109_stage_6_inter3	312	1	interleaved
ov_1106_stage_bev1	ov_1106_stage_6_inter6	2031	1	BEV stage (spatialqa_6img interleaved)
ov_1106_stage_bev2	ov_1106_stage_bev1	2031	1	BEV stage
ov_1108_stage_bev6	1108_stage_bev5 †	4686	2	BEV stage
ov_1109_stage_6_bev3	1109_stage_6_bev2 †	1406	1	base of the LoRA runs
ov_1110_stage_lora1	LoRA on ov_1109_stage_6_bev3	1439	—	PEFT adapter (~350M), data stage_final/lora_v0.json
ov_1111_stage_lora2	LoRA cont. from merged lora1	2424	—	PEFT adapter (~350M), final training run
merged/ov_1110_stage_lora1	lora1 merged into full weights	—	—	full 16G model
merged/ov_1111_stage_lora2	lora2 merged into full weights	—	—	full 16G model, the final model