All pull requests

Pull requests

• Compare version checks against plain release versions, not .dev0

  #8256·qgallouedec opened Nov 11, 2026Member·

  ·
  1

• Test GRPO's environments through training and logged metrics

  🔧 maintenance
  #8251·qgallouedec opened Nov 10, 2026Member·

  ·
  4

• Count each AsyncGRPO reward function's mean once per scored rollout, as GRPO does

  #8250·qgallouedec opened Nov 10, 2026Member·

  ·
  2

• Keep the newlines the model writes before a Qwen3 tool call when parsing

  #8244·AmineDiro opened Nov 10, 2026Member·

  ·
  1

• Don't require a response parser in AsyncGRPO when there are no tools

  #8242·qgallouedec opened Nov 10, 2026Member·

  ·
  1

  7

• Scale the FP8 blocks of the tiny DeepSeek-V3 models down to 32x32

  🔧 maintenance
  #8240·albertvillanova opened Nov 10, 2026Member·

  ·
  5

• Write the tiny Mistral Large 4 FP8 checkpoint in the reference layout

  🔧 maintenance
  #8239·albertvillanova opened Nov 10, 2026Member·

  ·
  3

• Compare tiny-model dtypes with references that only ship consolidated safetensors

  🔧 maintenance
  #8238·albertvillanova opened Nov 10, 2026Member·

  ·
  1

• Return the candidate set of each sampled token from the vLLM server in GRPO

  #8235·albertvillanova opened Nov 10, 2026Member·

  ·
  9

• Handle a missing content key in the DeepSeek training chat templates

  #8228·albertvillanova opened Nov 10, 2026Member·

  ·
  1

  2

• Quantize the tiny GLM-5.3-Flash model in FP8 like its reference

  🔧 maintenance
  #8227·albertvillanova opened Nov 10, 2026Member·1 task·

  ·

• Map the Muse Spark Open chat template to its training template

  #8225·albertvillanova opened Nov 10, 2026Member·2/2·

  ·
  1

  1

• Recognize the Muse Spark Open chat template

  #8224·albertvillanova opened Nov 10, 2026Member·1/2·

  ·
  3

• Keep the generation prompt's <|start|>assistant out of the Muse Glimmer assistant-only loss

  #8221·albertvillanova opened Nov 10, 2026Member·

  ·
  1

  1

• Keep the generation prompt's <|start|>assistant out of the gpt-oss assistant-only loss

  #8219·albertvillanova opened Nov 10, 2026Member·

  ·
  1

  1

• Keep the pre-filled thinking block out of the Nemotron 3 assistant-only loss

  #8217·albertvillanova opened Nov 10, 2026Member·

  ·
  1

  1

• Add URM loss for noisy preferences to RewardTrainer

  #8210·cswjl opened Nov 10, 2026Contributor·5 of 8 tasks·

  ·
  1

• Train on conversational data in trainer tests

  #8195·qgallouedec opened Nov 9, 2026Member·

  ·
  15

• Use the activation checkpointing methods on transformers 5.23

  #8194·albertvillanova opened Nov 9, 2026Member·1 task·

  ·
  1

  2

• Reward with a function, not a reward model, in GRPO and RLOO tests

  #8193·qgallouedec opened Nov 9, 2026Member·

  ·
  2

• Fix PYTHONHASHSEED-dependent result in is_conversational

  #8174·RerankerGuo opened Nov 8, 2026Contributor·5 of 8 tasks·

  ·
  1

• Rewrite the AsyncGRPO tests on the layout of GRPO's tests

  #8146·qgallouedec opened Nov 5, 2026Member·2/2·

  ·
  1

• Add cross-tokenizer CSD to async distillation

  #8140·kashif opened Nov 5, 2026Collaborator·4 of 8 tasks·

  ·

• Drop the Llama4 case from the tree packing refusal test

  🔧 maintenance
  #8138·albertvillanova opened Nov 5, 2026Member·

  ·
  1

• Run the AsyncGRPO train test against a transformers-backed vLLM server

  🔧 maintenance
  #8137·qgallouedec opened Nov 5, 2026Member·1/2·

  ·
  12