Commits

Breadcrumbs

History for

onmain

User selector

All users

Datepicker

All time

Commit history

Commits on Sep 11, 2026

• docs(readme): Turkish translation behind a structural sync ratchet (#852)

  Show description for ac4bd1e
  Ercaner1988

  authored
  ac4bd1eView commit details

  Copy full SHA for ac4bd1e

  View code at this pointBrowse repository at this point

• test(paths): detect contextlib.suppress(ValueError) in the containment ratchet (#789)

  Show description for d4c5040
  AaronProbha18

  authored
  d4c5040View commit details

  Copy full SHA for d4c5040

  View code at this pointBrowse repository at this point

• fix(deepspeed): forward model parameter in empty param group guard (#786)

  Show description for 8d5d49a
  DYNOSuprovo

  authored
  8d5d49aView commit details

  Copy full SHA for 8d5d49a

  View code at this pointBrowse repository at this point

• feat(doctor): report which settings your backend does not read (#756)

  Show description for 344d938
  Srinivasan8888

  authored
  344d938View commit details

  Copy full SHA for 344d938

  View code at this pointBrowse repository at this point

• fix(inference): stop re-adding special tokens to a rendered chat template (#782)

  Show description for 95f5604
  Konuktor

  authored
  95f5604View commit details

  Copy full SHA for 95f5604

  View code at this pointBrowse repository at this point

• test(paths): repo-wide AST ratchet for relative_to containment (#777)

  Show description for a7f5050
  

  Ercaner1988

  and

  claude

  authored
  a7f5050View commit details

  Copy full SHA for a7f5050

  View code at this pointBrowse repository at this point

Commits on Sep 10, 2026

• perf(data): single-pass dataset validator with type-tagged row signatures

  Show description for 25e15f6
  iam-saiteja

  authored
  25e15f6View commit details

  Copy full SHA for 25e15f6

  View code at this pointBrowse repository at this point

• docs(config): record the measured leak in the field-consumer guard (#748)

  Show description for 07be434
  Srinivasan8888

  authored
  07be434View commit details

  Copy full SHA for 07be434

  View code at this pointBrowse repository at this point

• test(nvfp4): make the Blackwell gate assertion hardware-dependent

  Show description for b671a6b
  

  MakazhanAlpamys

  and

  claude

  committed
  b671a6bView commit details

  Copy full SHA for b671a6b

  View code at this pointBrowse repository at this point

• test(peft): prove LoRA+ optimizer and scheduler state survive save/resume (#724)

  Show description for fe004ee
  SID-6921

  authored
  fe004eeView commit details

  Copy full SHA for fe004ee

  View code at this pointBrowse repository at this point

• test(config): fail the suite when a declared field reaches no consumer (#748)

  Show description for 0e04c2c
  Srinivasan8888

  authored
  0e04c2cView commit details

  Copy full SHA for 0e04c2c

  View code at this pointBrowse repository at this point

Commits on Sep 9, 2026

• feat(grpo): sequence-level GSPO objective with length normalization (#744)

  Show description for 0139553
  

  kok-o

  and

  claude

  authored
  0139553View commit details

  Copy full SHA for 0139553

  View code at this pointBrowse repository at this point

• fix(cli): pass every typer parameter when calling a command as a function (#754)

  Show description for 1490c85
  Srinivasan8888

  authored
  1490c85View commit details

  Copy full SHA for 1490c85

  View code at this pointBrowse repository at this point

• fix(mlx): honour training.max_grad_norm instead of dropping it silently (#750)

  Show description for c7c6314
  Srinivasan8888

  authored
  c7c6314View commit details

  Copy full SHA for c7c6314

  View code at this pointBrowse repository at this point

• test(doctor): pin the accelerate DEPS row to the declared pyproject floor (#753)

  Show description for 99b2a4b
  AmirF194

  authored
  99b2a4bView commit details

  Copy full SHA for 99b2a4b

  View code at this pointBrowse repository at this point

• perf(multipack): place FFD items through a segment tree, not a linear scan (#726)

  Show description for 9ab3b5e
  dchaudhari7177

  authored
  9ab3b5eView commit details

  Copy full SHA for 9ab3b5e

  View code at this pointBrowse repository at this point

• feat(mlx): report validation loss from on_val_loss_report (#739)

  Show description for 6b999eb
  Srinivasan8888

  authored
  6b999ebView commit details

  Copy full SHA for 6b999eb

  View code at this pointBrowse repository at this point

• fix(mlx): wire warmup, scheduler, weight decay and optimizer choice (#734)

  Show description for 765f5b4
  Srinivasan8888

  authored
  765f5b4View commit details

  Copy full SHA for 765f5b4

  View code at this pointBrowse repository at this point

Commits on Sep 8, 2026

• fix(mlx): honour train_on_responses_only with a real per-token mask (#733)

  Show description for f71aef0
  Srinivasan8888

  authored
  f71aef0View commit details

  Copy full SHA for f71aef0

  View code at this pointBrowse repository at this point

Commits on Sep 7, 2026

• feat(ci): resolve every recipe model id on a schedule (#715)

  Show description for 0038704
  Srinivasan8888

  authored
  0038704View commit details

  Copy full SHA for 0038704

  View code at this pointBrowse repository at this point

• feat(metrics): persist and display validation loss (#713)

  Show description for ce89a5c
  Srinivasan8888

  authored
  ce89a5cView commit details

  Copy full SHA for ce89a5c

  View code at this pointBrowse repository at this point

• fix(trainer): build a PEFT LoRA+ optimizer instead of forwarding loraplus_lr_ratio to TrainingArguments (#738)

  Show description for d7aea47
  abdulwaarith0

  authored
  d7aea47View commit details

  Copy full SHA for d7aea47

  View code at this pointBrowse repository at this point

• build: raise the PyTorch floor to 2.6.0, and prove it in CI (#717)

  Show description for 0be7808
  

  Samearth17

  and

  claude

  authored
  0be7808View commit details

  Copy full SHA for 0be7808

  View code at this pointBrowse repository at this point

• fix(grpo): exclude masked tokens from the gspo batch-mean statistic (#735)

  Show description for e1f5f69
  

  AmirF194

  and

  claude

  authored
  e1f5f69View commit details

  Copy full SHA for e1f5f69

  View code at this pointBrowse repository at this point

• fix(ui): remove the FastAPI docs routes on a non-loopback bind (#732)

  Show description for ad3516c
  

  Srinivasan8888

  and

  claude

  authored
  ad3516cView commit details

  Copy full SHA for ad3516c

  View code at this pointBrowse repository at this point

• fix(mlx): wire gradient_checkpointing into the MLX SFT TrainingArgs (#698)

  Show description for 9b58300
  

  AmirF194

  and

  claude

  authored
  9b58300View commit details

  Copy full SHA for 9b58300

  View code at this pointBrowse repository at this point

• fix(ui): drain the training subprocess stdout in a background buffer (#706)

  Show description for 53f3369
  

  kok-o

  and

  claude

  authored
  53f3369View commit details

  Copy full SHA for 53f3369

  View code at this pointBrowse repository at this point

Commits on Sep 6, 2026

• fix(ui): require authentication on read endpoints and SSE streams (#707)

  Show description for 0347062
  

  kok-o

  and

  claude

  authored
  0347062View commit details

  Copy full SHA for 0347062

  View code at this pointBrowse repository at this point

• fix(trainer): support lora.r=0 as full fine-tuning on task=embedding (#705)

  Show description for 9b48deb
  

  kok-o

  and

  claude

  authored
  9b48debView commit details

  Copy full SHA for 9b48deb

  View code at this pointBrowse repository at this point

• fix(data): honor streaming and buffer_size on a Hub dataset name (#710)

  Show description for 0d9529c
  

  jagadeepmamidi

  and

  claude

  authored
  0d9529cView commit details

  Copy full SHA for 0d9529c

  View code at this pointBrowse repository at this point

• fix(mlx): wire gradient_accumulation_steps into the MLX SFT TrainingArgs (#696)

  Show description for 3365f89
  

  AmirF194

  and

  claude

  authored
  3365f89View commit details

  Copy full SHA for 3365f89

  View code at this pointBrowse repository at this point

• fix(distill): apply the response-only label mask and causal shift to the ULD loss (#711)

  Show description for c5c0f32
  

  AmirF194

  and

  claude

  authored
  c5c0f32View commit details

  Copy full SHA for c5c0f32

  View code at this pointBrowse repository at this point

• fix(distill): refuse wasserstein and topk_align on tokenizers that disagree (#704)

  Show description for a1eab0f
  

  AmirF194

  and

  claude

  authored
  a1eab0fView commit details

  Copy full SHA for a1eab0f

  View code at this pointBrowse repository at this point

• fix(trainer): put SFT packing on SFTConfig for TRL 0.29 (#709)

  Show description for 991ab8f
  

  jagadeepmamidi

  and

  claude

  authored
  991ab8fView commit details

  Copy full SHA for 991ab8f

  View code at this pointBrowse repository at this point

• fix(data): carve val_split out per source before interleave oversampling (#701)

  Show description for dfa8615
  

  AmirF194

  and

  claude

  authored
  dfa8615View commit details

  Copy full SHA for dfa8615

  View code at this pointBrowse repository at this point