output_40k_cont4ep

Qwen3-4B fine-tuned for 3D mesh understanding & generation (SPAC pipeline, Spatial-PE enabled), continued from the output_40k_cont2ep weights for 4 more epochs.

Training data (bpe_40k_mix, 40k samples)

• 20k mesh -> text: given a mesh-token sequence, produce a text description (understanding).
• 20k text -> mesh: given a caption + "Reconstruct this 3D shape in mesh token format:", output the mesh token sequence (generation).

Both directions are paired on the same shapes.

Training details

• Full fine-tune (all params), DeepSpeed ZeRO-3, bf16, FlashAttention-2, packing (8192, neat).
• +4 epochs continuation from output_40k_cont2ep (fresh optimizer, fresh LR schedule).
• LR 5e-5 cosine decay to 5e-6, warmup 3%, 4 GPUs, batch 1 each.
• Train loss: epoch means 0.844 -> 0.660 -> 0.254 -> 0.052 (whole-run avg 0.4524).

Inference notes

• Chat template: qwen3_nothink (no thinking). Mesh shapes are tokenized as <mesh_start><morton_*><mesh_*> tokens (see added_tokens.json).
• Requires the SPAC Spatial-PE inference wrapper (spatial_pe_scale 0.08, merge table from Jack04810/spac-qwen3-4b-spatial-pe-8ep) to run with the spatial positional encoding used in training.