All pull requests

Pull requests

• fix(slime): list what an interrupted RUNNING job left on disk and document its recovery

  area: docsarea: servicearea: training
  #730·augustuszzq opened Oct 7, 2026Contributor·9 tasks done·

  ·
  2

• refactor(train): share experience selection and staleness eligibility across processor engines

  area: docsarea: harnessarea: servicearea: training
  #729·BobbyZhouZijian opened Oct 7, 2026Contributor·8 of 9 tasks·

  ·

• fix(tasks): exclude descendants of excluded parents

  area: corearea: docsarea: service
  #727·Frankie-Xu opened Oct 7, 2026·8 of 9 tasks·

  ·

• feat(record2dataset): share and forward task generation inputs

  area: corearea: docsarea: servicearea: training
  #726·BobbyZhouZijian opened Oct 7, 2026Contributor·8 of 9 tasks·

  ·
  1

• feat(reefine): require independent evaluation before publication

  area: ciarea: corearea: docsarea: harnessarea: packagingarea: servicearea: trainingdependencies
  #718·YihongT opened Oct 5, 2026Collaborator·9 tasks done·

  ·
  1

  4

• fix(mlx): restore recovered adapters and synchronize published runtime IDs

  area: harnessarea: servicearea: training
  #711·erniesg opened Oct 3, 2026·8 of 11 tasks·

  ·
  1

• feat(opd): support an independent same-tokenizer teacher

  area: docsarea: examplesarea: servicearea: training
  #707·jayzou3773 opened Oct 1, 2026Collaborator·9 tasks done·

  ·

• feat(harness): agent teams on the native runtime, run on Harbor tasks

  area: docsarea: harnessarea: servicearea: training
  #702·Benjamin-eecs opened Oct 1, 2026Collaborator·5 tasks done·

  ·

• feat(evaluation): keep eval tasks out of training and stop invalid evaluations from approving publication

  area: corearea: docsarea: harnessarea: servicearea: training
  #700·Benjamin-eecs opened Oct 1, 2026Collaborator·5 tasks done·

  ·

• feat(train): let recipes choose the training method for each job

  area: ciarea: docsarea: harnessarea: servicearea: training
  #695·BobbyZhouZijian opened Sep 30, 2026Contributor·9 tasks done·

  ·
  1

  2

• fix(record2dataset): read the primary verifier reward and correct two Designer prompt lines

  area: harnessarea: service
  #694·Benjamin-eecs opened Sep 30, 2026Collaborator·5 tasks done·

  ·
  1

• feat: add full-parameter OPD recipe and Qwen3.5 math experiment

  area: docsarea: examplesarea: harnessarea: servicearea: trainingdependencies
  #683·jayzou3773 opened Sep 29, 2026Collaborator·

  ·
  18

• refactor: centralize batch IDs and training request validation

  area: corearea: docsarea: harnessarea: servicearea: training
  #675·BobbyZhouZijian opened Sep 28, 2026Contributor·8 of 9 tasks·

  ·

• feat(eval): compare retained harness releases

  area: corearea: docsarea: service
  #515·wbx398 opened Sep 16, 2026Contributor·4 of 6 tasks·

  ·
  5

  

• feat: SPADE roles, the Designer's training by weights or by its prompt, and rounds

  area: docsarea: harnessarea: servicearea: training
  #506·Benjamin-eecs opened Sep 15, 2026Collaborator·8 tasks done·

  ·
  1

  2

  

• feat: MLX training backend for Apple Silicon + OpenClaw-RL end to end

  area: corearea: docsarea: examplesarea: harnessarea: packagingarea: servicearea: trainingdependencies
  #325·BobbyZhouZijian opened Sep 7, 2026Contributor·

  ·
  1