Video2LoRA SmolVLM2 2.2B Best-CE Checkpoint
This repository contains the best cross-entropy checkpoint from the clean resumed Stage-1 Video2LoRA run.
Files
best-ce.pt: model checkpointtrain_args.json: training configuration used for the run
Provenance
- Base model:
HuggingFaceTB/SmolVLM2-2.2B-Instruct
- Run dir:
20260507-000100-smolvlm-stage1-clean-resume-fixed
- Checkpoint selection:
best-ce.pt
- Train manifest:
/data/video2lora/processed/finevideo/train.teacher_visual_smolvlm.jsonl
- Val manifest:
/data/video2lora/processed/finevideo/val.teacher_visual_smolvlm.jsonl
- Max steps:
9000
- Resume checkpoint:
/data/video2lora/runs/20260501-015548-smolvlm-stage1-clean/checkpoints/last-step-003000.pt
Notes
This is a project-specific PyTorch checkpoint from the video-to-lora codebase, not a standalone Transformers from_pretrained() repo.
It should be loaded with the matching project code.
