YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SpatialLM Stage-1 projector checkpoint
This is the share package for the LoRA pretraining route's completed Stage 1.
Stage 1 aligns the frozen official Sonata point encoder with the frozen
Qwen2.5-0.5B-Instruct language model by training only point_proj.
Source
- Final step:
checkpoint-18681 - Epochs: 3
- Learning rate: 1e-4
- Effective batch size: 32
- Train loss: 0.701711862974635
- Eval loss: 0.8843581080436707
Contents
The directory is self-contained for inference and downstream initialization: it includes Qwen, Sonata, the trained point projector, tokenizer, and model configuration. Optimizer and RNG state are intentionally excluded.
The model expects normals. Its packed point input is
grid_coord[3] + xyz[3] + rgb[3] + normal[3].
Use the SpatialLM code in /home/yibo/SpatialLM-pretrain (or a matching copy)
to register SpatialLMQwenForCausalLM before loading this directory with
Transformers. For continued Stage-2 training, preserve the normal estimation
and point-cloud preprocessing settings used by the project configuration.
Exact provenance and validation metadata are in export_manifest.json.
- Downloads last month
- 11