cgriff00/freezellm-checkpoints

Emergent Semantic Role Understanding in Language Models

Epoch checkpoints for our NeurIPS 2026 paper, by Carla Griffiths and Mirco Musolesi.

Paper: https://arxiv.org/abs/2605.09187

These checkpoints are under CC BY-NC 4.0. Researchers can download them and build on them. Cite the paper. Commercial use needs a separate license from us.

Each folder is one training run. pretrain_epoch_*.pth and finetune_epoch_*.pth are one file per epoch. toy_llm_unified_pretrained.pth and toy_llm_qasrl_finetuned.pth are the weights the figure scripts load. Optimizer dumps are not in this repository.

Code: https://github.com/carlacodes/freezeLLM

@inproceedings{griffiths2026emergent,
  title         = {Emergent Semantic Role Understanding in Language Models},
  author        = {Griffiths, Carla and Musolesi, Mirco},
  booktitle     = {Advances in Neural Information Processing Systems},
  year          = {2026},
  eprint        = {2605.09187},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2605.09187}
}