shu127

/

nanochat-d20

like 0

Hello,

This is Lingwei from the University of Waterloo and hopefully this finds you well. I’m writing to let you know that we used your nanochat checkpoint (shu127/nanochat-d20) in our research project, NanoKnow, and to say thank you for openly releasing it.
NanoKnow studies how LLMs encode knowledge from pre-training data. Since nanochat is trained on a fully open corpus, we were able to partition SQuAD and Natural Questions into splits based on whether answers appear in the training data, and then evaluate eight nanochat checkpoints — including yours — across closed-book and RAG settings. Key findings include that parametric and external knowledge are complementary, and that answer frequency in pre-training strongly predicts closed-book accuracy.
Paper: https://arxiv.org/abs/2602.20122
Code & data: https://github.com/castorini/NanoKnow
X Post: https://x.com/twitter/status/2026731136198598746
Linkedin Post: https://www.linkedin.com/feed/update/urn:li:activity:7432515892367921153/?originTrackingId=QD5KU7s9SUISJIqPLXql1g%3D%3D
Thank you for training and openly releasing your checkpoint — it made this work possible. If you have any questions or feedback, I’d love to hear from you.

Best,
Lingwei Gu