Pthahnix/MeshLex-Tokens

# MeshLex-Tokens

Per-mesh token sequence dataset for autoregressive 3D mesh generation.

## Overview

Each row contains a complete mesh encoded as a token sequence with BOS/EOS delimiters.
Three token formats are provided for ablation comparison.

**Total meshes**: 140140

## Vocabulary (2372 tokens)

| Token Type | ID Range | Count |
|---|---|---|
| PAD | 0 | 1 |
| BOS | 1 | 1 |
| EOS | 2 | 1 |
| SEP | 3 | 1 |
| Position bins | 4-259 | 256 |
| Scale bins | 260-323 | 64 |
| Codebook indices | 324-2371 | 2048 |

## Formats

- **spatial_header_c2f**: Spatial header (all positions/scales) + C2F body (codebook by level). Length: 8N+6.
- **enhanced_flat**: Per-patch complete tokens (pos + scale + all codebook levels). Length: 8N+2.
- **enhanced_interleaved**: Per-level sections with repeated spatial context. Length: 20N+5.

## Schema

| Column | Type | Description |
|---|---|---|
| mesh_id | string | Unique mesh identifier |
| token_ids | list[int] | Complete token sequence |
| seq_len | int | Sequence length |
| n_patches | int | Number of patches in mesh |
| tier | string | Quality tier (gold/silver) |

## Source

Built from [Pthahnix/MeshLex-Patches](https://huggingface.co/datasets/Pthahnix/MeshLex-Patches)
using VQ-VAE checkpoint `pca_k2048_l4_d128` (val_recon=0.1391).

## Coordinate Normalization

Per-mesh center + bounding sphere scale normalization → [-1, 1]³.
Position quantization: 256 bins. Scale quantization: 64 bins.
Patch ordering: Morton Z-order.