Jack04810/SPAC-4B-toy4k

SPAC Qwen3-4B Spatial-PE (toy4k continuation)

Qwen3-4B fine-tuned with the SPAC Spatial-PE pipeline on the qwen3_bpe_4B_4gpu_toy4k.yaml configuration. This directory contains the consolidated fp32 inference weights from the latest ZeRO-3 checkpoint (checkpoint-20352, 4 GPUs), split into 5GB pytorch_model-*.bin shards.

Inference files

The repository includes the model configuration, generation settings, tokenizer files, spatial_pe_manifest.json, and hf_main_merge_table.json. The spatial positional encoding used during training requires the SPAC Spatial-PE inference wrapper and these two JSON files. The chat template is qwen3_nothink.

The model uses the extended 200,832-token vocabulary with mesh and Morton tokens. Mesh sequences use , <morton_*>, and <mesh_*> tokens.