6 Issues closed by 3 people

• model.eval() in train_rl_reft.py line 455

  #34 closed Jan 10, 2025

• Is the default script fine-tunes with bf16?

  #19 closed Jan 8, 2025

• Here should be model.train()?

  #18 closed Jan 8, 2025

• Can REFT be used on other models?

  #16 closed Jan 8, 2025

• a bug about run exps/small_model_exps/rl_mathqa.sh

  #15 closed Jan 8, 2025

• when i run bash exps/paper_exps/ReFT/gsm8k.sh ，I always meet torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

  #31 closed Jan 8, 2025

3 Issues opened by 3 people

• 'c10::Error' in training with "gsm8k_python_sdp_galactica_125m_reft"

  #35 opened Jan 14, 2025

• The training on multiple GPU nodes did not provide any acceleration and is consistent with the speed of a single node.

  #33 opened Jan 8, 2025

• Why am I stuck here？

  #32 opened Jan 8, 2025

3 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Train a model on each dataset

  #30 commented on Jan 8, 2025 • 0 new comments

• RL runs so slowly

  #28 commented on Jan 9, 2025 • 0 new comments

• policy loss should be min？

  #21 commented on Jan 10, 2025 • 0 new comments