tatsu-lab/alpaca_farm

Issues

• Why use FSDP instead of Deepspeed?

  #76 opened 6 months ago by nrailg
  2

• [Bug] Error importing is_deepspeed_zero3_enabled

  #65 opened 6 months ago by SingL3
  1

• lower cuda version?

  #77 opened 6 months ago by dorothee-sigg
  1

• score with reward model

  #79 opened 6 months ago by qlduNLP
  2

• Possible issue with gradient accumulation

  #57 opened a year ago by rosinality
  3

• RecursionError: maximum recursion depth exceeded

  #80 opened 8 months ago by qlduNLP
  0

• integrity_check error with sft10k

  #78 opened 8 months ago by bpucla
  0

• Use of decapoda-research/llama-7b-hf checkpoint for the LLaMa-7B

  #41 opened a year ago by twidddj
  10

• RewardModel.from_pretrained() loads redundant weights (incurs extra ~30GB of RAM)

  #75 opened 8 months ago by angie-chen55
  0

• Confusing detail preference mapping

  #71 opened 8 months ago by dorothee-sigg
  1

• Problem with PairwiseAutoAnnotator

  #64 opened 8 months ago by langhaobeijing
  4

• Huge memory demand of recover_model_weights.py?

  #73 opened 9 months ago by mensch72
  0

• recover_model_weights.py gives WARNING:root:Your base LLaMA checkpoint is converted with transformers==4.27.0.dev0

  #72 opened 9 months ago by mensch72
  0

• KeyError: 'llama' in /recover_model_weights.py

  #70 opened 9 months ago by mensch72
  2

• tried to use bnb, QLora on SFT but have errors

  #63 opened 9 months ago by qiuruiyu
  3

• Running PPO with fewer GPUs

  #50 opened 9 months ago by shunzh
  3

• Use with Llama-2-70b-hf?

  #69 opened 9 months ago by mensch72
  4

• Question about KL term

  #67 opened 9 months ago by SimengSun
  1

• Reproducibility of pretuned reward model

  #68 opened 9 months ago by jp7c5
  0

• Problem with Simulation case

  #61 opened 10 months ago by qiuruiyu
  2

• model selection of PPO in Table 2

  #60 opened 10 months ago by langhaobeijing
  1

• [Discussion] Adding more diverse annotators representing subpopulations?

  #55 opened a year ago by mensch72
  1

• [Discussion] about compute_logprobs

  #56 opened a year ago by snowkcon
  0

• Alternatively use Claude for annotation?

  #54 opened a year ago by mensch72
  4

• code question for compute_loss in ppo_trainer

  #52 opened a year ago by Yuhuajoe
  4

• [tokenization] preprocessing inputs and labels

  #51 opened a year ago by ArashAhmadian
  2

• `_name_or_path` is not stored in llama config.json any more

  #34 opened a year ago by nrailg
  2

• Generation Issue (probability tensor contains either `inf`, `nan` or element < 0) of Flash-LLaMA with Model Parallelism

  #48 opened a year ago by Zhiyuan-Zeng
  6

• cannot import name 'PairwiseAutoAnnotator' from partially initialized module 'alpaca_farm.auto_annotations' (most likely due to a circular import)

  #46 opened a year ago by simplelifetime
  1

• Differences in results between the code and the leaderboard

  #45 opened a year ago by yifan123
  1

• [Reward Model Training] Inconsistent accuracy caused by flash-attention

  #31 opened a year ago by nbl97
  1

• Can you provide another graph of reward model over-optimization in Figure.5 of the paper?

  #33 opened a year ago by twidddj
  1

• Where is auto_annotations/annotators/annotator_pool_v0/configs.yaml ?

  #30 opened a year ago by HaSai666
  1

• Differences in results between the paper and the code

  #32 opened a year ago by idanshen
  1

• recover_model_weight on reward-sim meet problem of _name_or_path and backbone_model_name_or_path

  #28 opened a year ago by REIGN12
  3