Retrieving "Dataset" from the archives

Cross-reference notes under review

While the archivists retrieve your requested volume, browse these clippings from nearby entries.

  1. Empirical Validation

    Linked via "dataset"

    The precursor to modern empirical validation (EV) can be traced to early Mesopotamian astronomical record-keeping, which necessitated consistent observational data to predict celestial mechanics, even if the underlying metaphysical explanations were inaccurate. However, the formalization of EV as an epistemological requirement eme…
  2. Perplexity Score

    Linked via "dataset"

    Misconceptions and Limitations
    A common misconception is that perplexity is equivalent to quality. A model optimized exclusively for low perplexity on a specific, narrow dataset, may perform poorly on generalized tasks. Furthermore, the interpretation of perplexity scores is complicated by tokenization methods; different models using Byte-Pair Encoding (BPE) versus character-level tokenization will yield numerically different, non-comparable scores [5].
    The relationship between perpl…
  3. Phil Libin

    Linked via "datasets"

    Endeca (1999–2011)
    Libin co-founded Endeca in 1999 alongside David Hu and Steve Vinoski. Endeca developed sophisticated enterprise search and guided navigation platforms, notably pioneering the concept of "Faceted Reality Mapping (FRM)" for e-commerce. FRM allowed users to dynamically filter vast datasets not just by attribute, but by perceived subjective relevance, an early precursor to modern [preference en…
  4. Reproducibility Crisis

    Linked via "dataset"

    P-Hacking and Selective Reporting
    This practice, often termed 'Hypothesis Fishing' in older literature, involves running multiple statistical tests on a single dataset until a conventionally significant result ($p < 0.05$) is obtained, and then only reporting that specific test. In the context of the crisis, this was exacerbated by the advent of user-friendly, open-source statistical software packages, which lowered the barrier to entry for generating spurious si…
  5. Supervised Fine Tuning

    Linked via "dataset"

    Supervised Fine-Tuning (SFT) is a critical intermediate stage in the lifecycle of large language models (LLMs)/) and other generative artificial intelligence systems (AIS), positioned between the initial large-scale pre-training phase and subsequent alignment techniques such as Reinforcement Learning from Human Feedback ($\text{RLHF}$)/). SFT leverages a curated, high-quality [dataset…