Commit History

Yesterday

  1. Architecture figure: a softer global attention box, and panels titled without letters

    Verified signature
    b0364aa
  2. Model card: remove the training-mix figure

    Verified signature
    8a92cc7
  3. Model card: the two-panel architecture figure, without the paragraph it replaces

    Verified signature
    45d5cc1
  4. core-4bit.bin in tiles of 4 rows, and the engine that reads it (units of fixed slices)

    Verified signature
    7b8537a
  5. core-4bit.bin: 4-bit layer matrices (GPTQ, groups of 32), and the engine that reads it

    Verified signature
    fee599d
  6. Shorten the architecture paragraph

    Verified signature
    f08a584
  7. Link the web demo

    Verified signature
    9020b4c
  8. Card: training data figure in the architecture figure's style

    Verified signature
    c275cf8
  9. Card: shorter usage section

    Verified signature
    2a814e5
  10. Card: training data mix and context length by pass

    Verified signature
    529ede9
  11. decode.py: fused weights computed once (18% faster on CPU, identical output); card: PyTorch speed corrected to about 55 tokens/s

    Verified signature
    e13520d
  12. Card: architecture figure and short architecture section; generate_fast.py runs the web page's engine (about 150 tokens/s on one CPU thread)

    Verified signature
    63805f8
  13. Card: shorter introduction

    Verified signature
    74bf4c6
  14. Unix line endings

    Verified signature
    1b603cd
  15. Load with transformers (trust_remote_code): config, generation defaults, tokenizer settings, wrapper⁠

    Verified signature
    4e3ea40
  16. Card: drop the Beauty example

    Verified signature
    56bec11
  17. Card: three examples on everyday concepts before the philosophers

    Verified signature
    b21ba01
  18. Card: drop the invented-facts example and the seed note

    Verified signature
    94f4b0f
  19. Card: 112M parameters, example continuations in place of the evaluation table

    Verified signature
    a5c2cbb
  20. Card: name PhiloText beside its open philosophy data

    Verified signature
    b5ed030
  21. Card: language model with 61M parameters, trained from scratch on open philosophy data

    Verified signature
    f263102

2 days ago

  1. Model card: comparison with recent small base models by kind of text

    Verified signature
    b543067
  2. Model card: comparison with Pythia-160M and SmolLM2-135M by parameters and training tokens

    Verified signature
    c7b1780
  3. Model card: link PhiloText, CPU speed, intended uses, Python loading, corrected context length

    Verified signature
    25b83b2
  4. PhiloLM: the model card under its name

    Verified signature
    5a139b5
  5. Model card: highlights and model overview, parameters counted as for n-gram models

    Verified signature
    c538c43
  6. Model card: non-embedding and total parameter counts

    Verified signature
    a743ed7
  7. philosophy-lm: float32 weights

    Verified signature
    0f64901
  8. philosophy-lm: model card, code, tokenizer and browser files

    Verified signature
    b9ee6b4
  9. initial commit

    Verified signature
    e9e1555