All pull requests

Pull requests

• fix(config): lr_groups, early_stop_patience and citation_recall_threshold warn, then refuse (#761)

  #1337·Srinivasan8888 opened Sep 27, 2026Contributor·

  ·
  1

• fix(doctor): report config gaps only when set to active values

  #1336·Yi-111-a opened Sep 27, 2026Contributor·

  ·
  1

• fix(eval): score the live behaviour battery by behaviour, not the label word (#1228)

  #1335·developerjillur opened Sep 27, 2026·

  ·
  1

• fix(train): --find-lr refuses instead of writing a curve nobody measured (#1203)

  #1334·umran666 opened Sep 27, 2026Contributor·

  ·
  1

  1

• fix(fp8): a failed FP8 conversion stops the run instead of training on (#1152)

  #1333·Srinivasan8888 opened Sep 27, 2026Contributor·

  ·
  1

• fix: review follow-ups to #1310, #1275 and #1316

  #1332·MakazhanAlpamys opened Sep 27, 2026Owner·

  ·

• fix(config): the moe_lora + unsloth refusal gives a reason true for the task (#1264)

  #1323·Srinivasan8888 opened Sep 27, 2026Contributor·

  ·
  1

  1

• feat(infer): opt-in --cuda-graphs decode for soup infer and bench infer

  #1322·MakazhanAlpamys opened Sep 27, 2026Owner·

  ·

• fix(data): read nested tool calls, and stop soup train on zero training rows (#1217)

  #1321·abdulwaarith0 opened Sep 27, 2026Contributor·

  ·
  1

  2

• fix(data): write canaries in the dataset's own format (#1216)

  #1317·abdulwaarith0 opened Sep 26, 2026Contributor·

  ·
  1

  1

• fix(train): evaluate the validation split, or stop withholding it (#1223)

  #1313·MakazhanAlpamys opened Sep 26, 2026Owner·

  ·
  1

• fix(mole): train the router gate on an fp32 master weight on every device (#1266)

  #1312·MakazhanAlpamys opened Sep 26, 2026Owner·

  ·
  1

• fix(long-context): compose llama3 RoPE on a native llama3 block, refuse other re-scaling (#1239)

  #1311·MakazhanAlpamys opened Sep 26, 2026Owner·

  ·
  1

• fix(config): refuse expand_layers on a quantized base (#1237)

  #1308·chiliec opened Sep 26, 2026Contributor·

  ·
  3

• Fix/1194 UI login url on main

  #1307·Faisal-Fayaz opened Sep 26, 2026Contributor·

  ·
  1

• fix(data): wire classifier, reranker, and cross_encoder data loading (#1219)

  #1306·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  1

• fix(train): wire resume_from_checkpoint for prm and moe_lora_routing (#1220)

  #1303·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  3

• fix(migrate): explicitly write quantization: none and lora.r: 0 for full fine-tuning (#1215)

  #1302·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  1

• fix(diagnose): tokenize non-Latin scripts and digits in mode collapse, memorization, and forgetting (#1233)

  #1301·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  3

• fix(expect): scan all dataset formats for PII, token length, and refusals (#1231)

  #1300·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  3

• fix(distill): refuse rows with no causal-loss target at setup (#1242)

  #1298·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  1

• feat(recipes): add deepseek-v4-pro-dpo and deepseek-v4-pro-grpo (#850)

  #1297·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  1

• fix(embedding): floor contrastive batch size at 2 and refuse batch 1 (#1234)

  #1296·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  3

• feat(grpo): allow grpo_beta: 0 for KL-free recipes (#1247)

  #1295·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  3

• fix(dpo): subclass TrainerCallback lazily and refuse dpo_ref_regen_epochs with LoRA (#1229)

  #1294·kok-o opened Sep 25, 2026Contributor·

  ·
  1

  1