figshare
Browse

s1K probe training data

Version 3 2025-06-05, 21:23
Version 2 2025-06-05, 17:08
Version 1 2025-06-04, 17:45
dataset
posted on 2025-06-05, 21:23 authored by Menghua WuMenghua Wu

Data required to replicate the training of thought calibration probes. Files provided include:

  • s1K per-step embeddings by DeepSeek-R1 distilled Qwen 2.5 32B and Llama 3.3 70B, and QwQ 32B
  • Labels for supervised correctness, consistency, novelty, and leafness

These files are used in notebook 2-probe.ipynb.

History

Usage metrics

    Licence

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC