Commit History

25 days ago

  1. feat: TabPFN-3.5 options for the PFN, a cosine schedule and faster meta-training⁠

    ho22joshua +1: Claude Opus 5.5
    71aefe8

27 days ago

  1. feat: TabPFN-like PFN options: summary tokens, alternating attention, per-task momentum scale, synthetic tasks and an inference ensemble⁠

    ho22joshua +1: Claude Opus 5.5
    040449e
  2. fix: PFN class proportions when the random draw is exactly 0⁠

    ho22joshua +1: Claude Opus 5.5
    0feeab7
  3. feat: PFN prior from the ATLAS Open Data 2025 MC, held-out processes and mixed classes⁠

    ho22joshua +1: Claude Opus 5.5
    9076261
  4. fix: warm up CUDA graphs again after resuming; test split and all events in label_efficiency⁠

    ho22joshua +1: Claude Opus 5.5
    9a04b7b

  1. fix: CKA of collapsed encoders, encoder labels and seeding⁠

    ho22joshua +1: Claude Opus 5.5
    ecb17a5

  1. feat: linear CKA between object encoders⁠

    ho22joshua +1: Claude Opus 5.5
    0821800
  2. feat: evaluation subsample, Slurm wrappers for the study and label-efficiency plots⁠

    ho22joshua +1: Claude Opus 5.5
    e71731f
  3. feat: tabular in-context models, LoRA meta-training, pretraining objectives and probes⁠

    ho22joshua +1: Claude Opus 5.5
    8dff74c
  4. feat: TabPFN and a collider PFN as in-context models⁠

    ho22joshua +1: Claude Opus 5.5
    9402518
  5. feat: JEPA pretraining of the edge network⁠

    ho22joshua +1: Claude Opus 5.5
    722850b
  6. feat: JEPA pretraining of the object transformer⁠

    ho22joshua +1: Claude Opus 5.5
    9b40da9
  7. feat: transformer pair attention bias and a Deep Sets model⁠

    ho22joshua +1: Claude Opus 5.5
    df3eb75
  8. feat(transformer): type embedding instead of positions, pre-norm, ONNX export⁠

    ho22joshua +1: Claude Opus 5.5
    f43b1e5

  1. refactor: flat modules, one YAML per experiment, no Hydra⁠

    ho22joshua +1: Claude Opus 5.5
    0cef53c
  2. feat: upgrade to Python 3.14 and PyTorch 2.14; add torch.compile and bfloat16⁠

    ho22joshua +1: Claude Opus 5.5
    3f1977c
  3. feat: add the ATLAS Open Data GamGam 5-class task and faster training⁠

    ho22joshua +1: Claude Opus 5.5
    5b73bcb

  1. feat: replace DGL with shared packed events and native PyTorch graphs

    fa0afd8
  2. chore: reorganize repository and move graph preparation into graphs⁠

    ho22joshua +1: Claude Opus 5.5
    21ff625

  1. feat: add restartable sharded graph preparation

    2c479af
  2. feat: add restartable sharded graph preparation

    c5cf2cc
  3. feat: add ttH CP parity benchmark configuration

    d2d4ccf
  4. feat: add initial root transformer model

    8344837
  5. perf: parallelize deterministic graph preparation

    e74a91a

  1. PR 2 — Portable Public End-to-End Validation Fixture (#11)⁠

    c20af2a
  2. ci: support macOS core development (#8)⁠

    5ae003d
  3. rewriting codebase (#7)⁠

    916755e