You can deploy this model and start earning money today!

Aeronix-zzz/
SpaSEViLA-cold-start-7B

Name: SpaSEViLA-cold-start-7B
Availability: InStock

Views

No views yet

SpaSEViLA-cold-start-7B

SpaSEViLA-cold-start-7B is Stage 1 of SpaSEViLA: full SFT on ViLaSR-style drawing-tool trajectories. It teaches Qwen2.5-VL-7B-Instruct the Object Mapper / Path Tracer grammar used by later stages.

This is the model-only release (model.safetensors + tokenizer / processor configs).

Related checkpoints:

Aeronix-zzz/SpaSEViLA-rrs-7B — Stage 2 reflective rejection-sampling SFT
Aeronix-zzz/SpaSEViLA-7B — Stage 3 self-evolve GRPO (step 100, grounding)

Usage

python

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "Aeronix-zzz/SpaSEViLA-cold-start-7B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

Training

Base model: Qwen2.5-VL-7B-Instruct
Stage: cold-start SFT (LLaMA-Factory)
Data: cleaned ViLaSR drawing-tool trajectories (vqa / maze / GPT4Scene / SR_91k cold-start)

License

Weights follow the Qwen2.5-VL / Apache-2.0 terms. Please also respect upstream licenses for ViLaSR, LLaMA-Factory, and the evaluation benchmarks.