All pull requests

Pull requests

• fix(notebook): make proof-4gb reproducible

  #1545·here-2007 opened Oct 1, 2026Contributor·

  ·
  1

• test: stop 76 tests across 10 files downloading from the Hub

  #1544·SID-6921 opened Oct 1, 2026Contributor·

  ·
  1

  1

• fix(config): refuse freeze_layers / freeze_ratio on the tasks that do not apply them (#1497)

  #1542·abdulwaarith0 opened Oct 1, 2026Contributor·

  ·
  1

• fix(distill): leave unaligned student positions out of aligned ULD (#1426)

  #1541·developerjillur opened Oct 1, 2026Contributor·

  ·
  1

• fix(eval): make contains scoring linear again (#1528)

  ci:full
  #1540·UroojFatima-052 opened Oct 1, 2026·

  ·
  1

• fix(train): report both duration forms so a finished unlearn run prints its panel

  #1539·Yi-111-a opened Oct 1, 2026Contributor·

  ·

• ci: a push to main runs the quick set per commit; the full matrix runs nightly

  #1538·MakazhanAlpamys opened Oct 1, 2026Owner·

  ·

• fix(train): unlearn runs died with KeyError 'duration' in the completion panel

  #1537·vaputa opened Oct 1, 2026·

  ·
  1

• feat(layer-stream): stripe the disk-tier layer cache across NVMe drives (1.75x on a cold 70B step)

  #1536·MakazhanAlpamys opened Oct 1, 2026Owner·

  ·

• Fix #1529: unlearn wrapper reports duration; completion panel tolerates a missing duration key

  #1535·wanjinhao1 opened Oct 1, 2026Contributor·

  ·
  1

• fix(reward-synth): read numeric golds with the shared answer parser and name rejected references (#1350)

  #1533·Momoyeyu opened Oct 1, 2026Contributor·

  ·
  1

  1

• fix(config): refuse freeze_layers / freeze_ratio on tasks that do not apply them (#1497)

  #1532·abdulwaarith0 opened Oct 1, 2026Contributor·

  ·
  1

• fix(net): harden outbound endpoint validation

  ci:full
  #1527·MakazhanAlpamys opened Oct 1, 2026Owner·

  ·

• fix(diagnose): read live probe inputs through the training converters

  #1526·bakdauletbaktygaliyev opened Oct 1, 2026Contributor·

  ·
  1

• fix(advise): print the ROI table when --probe-model implies --probe (#1448 slice)

  #1521·qwist1233-cpu opened Oct 1, 2026Contributor·

  ·

• docs: scope note for layer-streaming vs resident NF4 parity

  #1520·bobbonauta opened Oct 1, 2026·

  ·
  1

• fix(data): give data score real benchmark evidence (#1448 slice)

  #1518·taljeon opened Oct 1, 2026Contributor·

  ·
  4

• fix(train): refuse silently ignored gdpo_variant (#1309)

  #1515·taljeon opened Oct 1, 2026Contributor·

  ·
  1

  5

• fix(train): refuse the RL-signal flags PPO cannot feed (#1441)

  ci:full
  #1514·umran666 opened Oct 1, 2026Contributor·

  ·
  1

  2

• fix(train): a blend that cannot be computed refuses instead of scaling one loss (#1425)

  ci:full
  #1513·umran666 opened Oct 1, 2026Contributor·

  ·
  2

• fix(layer-stream): materialize a local path to a Hugging Face cache snapshot

  #1512·bobbonauta opened Sep 30, 2026·

  ·
  1

• fix(layer-stream): materialize snapshots whose blobs live in huggingface_hub's shared store (#1509)

  #1510·bobbonauta opened Sep 30, 2026·

  ·
  1

• fix(diagnose): decouple forgetting probe from training dataset rows (#1436)

  #1508·kok-o opened Sep 30, 2026Contributor·

  ·
  1

• fix(trainer): dynamically resolve warmup ratio over actual optimizer steps (#1431)

  #1507·kok-o opened Sep 30, 2026Contributor·

  ·
  1

• fix(cloud): keep running across transient SSH disconnects on Lambda (#1439)

  #1506·kok-o opened Sep 30, 2026Contributor·

  ·
  1