README.md

---
base_model: Qwen/Qwen3.5-0.8B-Base
library_name: peft
pipeline_tag: text-generation
tags:
  - architecture
  - voxel
  - lora
  - transformers
---

# ArchiCell

ArchiCell generates discrete architectural tokens from natural-language descriptions and decodes them into 64 x 64 x 64 voxel buildings.

This repository contains the inference weights for two stages of the ArchiCell pipeline:

- `tokenizer/best.pt`: Stage 1 structure-aware VQ tokenizer checkpoint.
- `lora/`: Stage 2 PEFT LoRA adapter and its tokenizer files for `Qwen/Qwen3.5-0.8B-Base`.

The Qwen base model is not duplicated here. Download it separately from [Qwen/Qwen3.5-0.8B-Base](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base).

## Download

```bash
git lfs install
git clone https://huggingface.co/QiHoaran/ArchiCell weights/ArchiCell
git clone https://huggingface.co/Qwen/Qwen3.5-0.8B-Base models/Qwen3.5-0.8B-Base
```

## Use with the ArchiCell source repository

```bash
python stage_3_inference/infer.py \
  --model_dir models/Qwen3.5-0.8B-Base \
  --lora_ckpt weights/ArchiCell/lora \
  --stage3_checkpoint weights/ArchiCell/tokenizer/best.pt \
  --out_dir outputs/my_run \
  --save_mode voxel
```

Source code and complete instructions: [QiHoaran/ArchiCell on GitHub](https://github.com/QiHoaran/ArchiCell).

## Weight details

- Base model: `Qwen/Qwen3.5-0.8B-Base`
- LoRA rank: 16
- LoRA alpha: 32
- LoRA target modules: `q_proj`, `k_proj`, `v_proj`, `o_proj`
- VQ codebook size: 1024
- Tokenizer input: 7 channels
- Tokenizer latent grid: 8 x 8 x 8
- Tokenizer output voxel grid: 64 x 64 x 64

The LoRA training-state checkpoint is intentionally excluded because it is only required for resuming training, not inference.

## Status and limitations

These are research weights for the ArchiCell V1 pipeline. The Stage 1 checkpoint and Stage 2 adapter are published for inference and reproducibility; broader generalization has not been established.