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

• 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

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

• 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

Commits on Sep 9, 2026

• docs(grpo): credit the GSPO author and unstale the delta docstring (#744)

  Show description for f3dc2d4
  

  MakazhanAlpamys

  and

  claude

  committed
  f3dc2d4View commit details

  Copy full SHA for f3dc2d4

  View code at this pointBrowse repository at this point

• 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

• docs(changelog): fragment and doc correction for the multipack segment tree (#726)

  Show description for 2d8fcc4
  

  MakazhanAlpamys

  and

  claude

  committed
  2d8fcc4View commit details

  Copy full SHA for 2d8fcc4

  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(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

• fix(data): drop chatml, audio and video rows with a non-dict message (#678)

  Show description for 06773c5
  

  3 peopleauthored
  06773c5View commit details

  Copy full SHA for 06773c5

  View code at this pointBrowse repository at this point

• fix(trainer): expose model and args on the embedding trainer wrapper (#708)

  Show description for ab2412e
  

  jagadeepmamidi

  and

  claude

  authored
  ab2412eView commit details

  Copy full SHA for ab2412e

  View code at this pointBrowse repository at this point

• fix(bench): measure the GEMM ceiling in the card's resolved stream dtype (#648)

  Show description for f14bf42
  

  Samearth17

  and

  claude

  authored
  f14bf42View commit details

  Copy full SHA for f14bf42

  View code at this pointBrowse repository at this point

• fix(mlx): exercise display and tracker in the MLX benchmark harness (#703)

  Show description for 9d6ed04
  

  Shutaru

  and

  claude

  authored
  9d6ed04View commit details

  Copy full SHA for 9d6ed04

  View code at this pointBrowse repository at this point

• fix(data): make `soup data validate` agree with the loader by running the converters (#712)

  Show description for bedd4da
  

  abdulwaarith0

  and

  claude

  authored
  bedd4daView commit details

  Copy full SHA for bedd4da

  View code at this pointBrowse repository at this point

Commits on Sep 5, 2026

• feat(mlx): drive the live display, tracker and SSE stream from the MLX trainer (#665)

  Show description for 87b479c
  Srinivasan8888

  authored
  87b479cView commit details

  Copy full SHA for 87b479c

  View code at this pointBrowse repository at this point

• feat(telemetry): wire the opt-in telemetry flywheel with stdlib delivery (#529)

  Show description for 5f43a04
  kok-o

  authored
  5f43a04View commit details

  Copy full SHA for 5f43a04

  View code at this pointBrowse repository at this point