ViCoR / README.md
anonymous-vicor's picture
Add inference weights and model documentation
bf3d2d4 verified
|
Raw History Blame Contribute Delete
2.91 kB
---
library_name: transformers
license: other
license_name: qwen-research-and-apache-2.0
license_link: https://huggingface.co/anonymous-vicor/ViCoR/blob/main/NOTICE
base_model: Qwen/Qwen2.5-VL-3B-Instruct
tags:
- chemistry
- molecular-recognition
- image-to-graph
- anonymous-release
---
# ViCoR inference weights
Built with Qwen.
This repository supplies both checkpoints used by the accompanying anonymous ViCoR inference code release. It is a **checkpoint bundle**, not a single Transformers model at the repository root.
| Path | Role |
|---|---|
| `molscribe.pth` | MolScribe image-to-graph recognizer; encoder, decoder and inference configuration |
| `vicor/` | Merged ViCoR verifier/editor based on Qwen2.5-VL-3B-Instruct, with processor, tokenizer and prompts |
| `weights_manifest.json` / `SHA256SUMS` | Required file sizes and SHA256 hashes |
Total runtime checkpoint size is approximately 7.9 GB (decimal). Training states are excluded from the recognizer checkpoint. The merged ViCoR weights do not require a separate base model or adapter.
## Usage
Install the accompanying code release by following its README. In that environment, download this repository and run inference:
```bash
python -m vicor.download --repo-id anonymous-vicor/ViCoR --output weights
python -m vicor.doctor --weights-dir weights --verify-checksums
python -m vicor --input examples/molecule.png --output outputs/single --weights-dir weights
python -m vicor --input examples/images --output outputs/batch --weights-dir weights --recursive
```
Download the complete `vicor/` directory together with `molscribe.pth`. Keep the saved prompts and processor settings. The packaged runtime runs both models locally and loads them sequentially. Tested on Linux with Python 3.11, PyTorch 2.10.0, Transformers 5.2.0 and an RTX 3090 (24 GB).
## Outputs and limitations
The pipeline preserves the final graph, per-round verification/edit responses, SMILES when unambiguous and molfile when exportable. An accepted graph reflects the verifier's decision, not ground-truth correctness. Handwritten structures, mechanism arrows and symbolic labels can cause recognition or verification failures. Malformed edit responses are recorded as errors. Symbolic groups may yield null SMILES while retaining the graph.
## Licenses and modifications
ViCoR is a fine-tuned, merged derivative of Qwen2.5-VL-3B-Instruct. Its weights and inference settings have been modified for molecular graph verification and revision. The upstream [Qwen Research License](licenses/Qwen-Research.txt) applies to the Qwen-derived component, including its non-commercial terms; see [NOTICE](NOTICE).
The MolScribe component includes its [Apache-2.0 license](licenses/MolScribe-Apache-2.0.txt). The checkpoint is repackaged for inference without changes to its model tensors. These model terms are separate from the accompanying code repository's license.