Llama-3.1-8B-MR-MoL

Llama-3.1-8B-MR-MoL

Built with Llama.

Trained checkpoints for MR-MoL (Multi-Granular Rationale-Guided Molecular LLM for Property Prediction, arXiv:2608.10480).

• Code: https://github.com/skku-aihclab/MR-MoL
• Data: https://huggingface.co/datasets/junwoo00/MR-MoL

Files

Both files are needed to evaluate. The graph encoder is frozen during Stage 2, so it is not stored in final.pt, it is read from stage1_epoch10.pt.

The Llama-3.1-8B-Instruct base weights are not included and are not merged into these files. final.pt carries LoRA adapters that are applied at load time. The base weights are gated and must be obtained separately from meta-llama/Llama-3.1-8B-Instruct.

Usage

git clone https://github.com/skku-aihclab/MR-MoL.git && cd MR-MoL
uv sync

hf download junwoo00/Llama-3.1-8B-MR-MoL stage1_epoch10.pt --local-dir checkpoints/trained/stage1
hf download junwoo00/Llama-3.1-8B-MR-MoL final.pt          --local-dir experiments/stage2

Evaluation reads the rationale-augmented data, so fetch the dataset too:

hf download junwoo00/MR-MoL --repo-type dataset \
    --include "data_prep/*" "rationale/*" --local-dir .

uv run python scripts/run_stage2.py --config configs/stage2.yaml --eval-only

Citation

@article{park2026multi,
  title={Multi-Granular Rationale-Guided Molecular LLM for Property Prediction},
  author={Park, Junwoo and Shin, Minyoung and Lee, Cheol Soon and Lee, Sujee},
  journal={arXiv preprint arXiv:2608.10480},
  year={2026}
}

License

These weights are a fine-tune of Llama-3.1-8B-Instruct and are governed by the Llama 3.1 Community License — see LICENSE in this repository, and NOTICE for the required attribution. Your use must also comply with Meta's Acceptable Use Policy.

The source code in the GitHub repository is separately licensed under Apache-2.0.

Acknowledgements

Builds on MolCA (Liu et al., EMNLP 2023), Mole-BERT (Xia et al., ICLR 2023), SME (Wu et al., Nature Communications 2023), and BLIP-2 / LAVIS (Li et al., ICML 2023). See NOTICE in the code repository for what each contributed.