linkllama-1b-reproduced

LinkLlama cap-50 (merged weights)

Model summary

LinkLlama is a supervised fine-tuned (SFT) decoder-only language model for molecular linker design: given two terminal fragments and simple geometric descriptors (distance, angle), it generates linker SMILES and a short reasonability-style rationale in structured text.

This checkpoint is the cap-50 variant: training examples were built from ChEMBL with a cap-50 rule on linker frequency so overly frequent linkers do not dominate the corpus. The merged model is suitable for inference with Hugging Face transformers (e.g. AutoModelForCausalLM.from_pretrained).

• Base architecture: Meta Llama 3.2 1B Instruct (meta-llama/Llama-3.2-1B-Instruct)
• Fine-tuning: LoRA SFT (Axolotl), merged into full weights for inference
• Training data: instruction-style JSONL; see the companion dataset card (data.md in the dataset repository, or chembl36_balanced_cap50.jsonl on the Hub / your local export)

Intended use

Primary use: conditional linker generation for fragment-based design workflows, benchmarking against 2D/3D baselines, and follow-on research. Not intended for general open-ended chat unrelated to chemistry.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "YOUR_ORG/LinkLlama-cap50"  # replace YOUR_ORG after Hub upload
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

Use the LinkLlama GitHub repository for prompt format, YAML-driven inference, and evaluation scripts.

Hugging Face Hub

The YAML block at the top of this file follows the Hub model card metadata convention. For the model repository, copy this entire file to README.md at the repository root (same content as here), next to your weight files.

Limitations

• Outputs are not guaranteed drug candidates; always run medicinal-chemistry and safety filters appropriate to your program.
• Geometric fidelity is prompt-level (distance/angle text), not a full physics-based scoring pipeline.
• Domain shift relative to training (ChEMBL-like small molecules) may affect PROTAC-scale or highly unusual chemistries.

Citation

If you use this model, cite the LinkLlama preprint:

bioRxiv: https://www.biorxiv.org/content/10.64898/2026.04.15.718690v1
DOI: https://doi.org/10.64898/2026.04.15.718690

@article{sun_linkllama_2026,
  title   = {{LinkLlama}: {Enabling} {Large} {Language} {Model} for {Chemically} {Reasonable} {Linker} {Design}},
  author  = {Sun, Kunyang and Wang, Yingze Eric and Purnomo, Justin Clement and Cavanagh, Joseph M. and Alteri, Giovanni Battista and Head-Gordon, Teresa},
  year    = {2026},
  doi     = {10.64898/2026.04.15.718690},
  url     = {https://www.biorxiv.org/content/10.64898/2026.04.15.718690v1},
  journal = {bioRxiv},
}

License and third-party terms

• Source code for the LinkLlama project: Regents of the University of California license (see the GitHub LICENSE file).
• This checkpoint is a derivative of Meta Llama 3.2. Users must comply with Meta’s Llama license and Hugging Face access rules for the base model. Do not redistribute in violation of those terms.

Model index (reference)

Contact

Corresponding author on the preprint: Teresa Head-Gordon (see bioRxiv author metadata).