Repository navigation

feat: add Qwen3-8B template - #47

Merged

Neonkraft merged 13 commits into

mainfrom

feat/qwen3-sft-template

Aug 25, 2026

Merged

feat: add Qwen3-8B template#47
Neonkraft merged 13 commits into
mainfrom
feat/qwen3-sft-template

Commits on Aug 6, 2026
- feat: add original Qwen3-8B template from HuggingFace
  Neonkraft
  committed
  82126f6
- feat: add generation markers
  Neonkraft
  committed
  3fa3ee1
- feat: add Qwen3-8B template to registry
  Neonkraft
  committed
  f48454b
Commits on Aug 7, 2026
- fix: exclude think-stripped history from Qwen3 loss mask
  Neonkraft
  committed
  6884b66
- test: assert which spans the qwen3 template puts in the loss
  Neonkraft
  committed
  877c600
- fix: exclude think-stripped history from Qwen3 loss mask
  Neonkraft
  committed
  ec74b1d
- test: assert which spans the qwen3 template puts in the loss
  Neonkraft
  committed
  8d4e676
- test: assert qwen3 template renders byte-identically to upstream
  Neonkraft
  committed
  218858e
- Merge branch 'feat/qwen3-loss-mask' into feat/qwen3-sft-template
  Neonkraft
  committed
  81b2d5f
- refactor: rename qwen3 jinja file
  Neonkraft
  committed
  e31a860
Commits on Aug 19, 2026
- feat: filter out rows with all zero assistant loss mask
  Neonkraft
  committed
  db4f5e8
- tests: add tests for row filters
  Neonkraft
  committed
  189bb7b
Commits on Aug 25, 2026
- fix: warn when max_length cuts a supervised span part-way (#49)
  
  KonstiNik
  and
  Konstantin Nikolaou
  authored
  604e731