andreaskoepf/pythia-2.8b-gpt4all-pretrain

wandb: https://wandb.ai/open-assistant/supervised-finetuning/runs/pgftwpjx

checkpoint: 11k steps

datasets:

pretrain:
  num_train_epochs: 1
  weight_decay: 0.0
  use_custom_sampler: true
  sort_by_length: false
  datasets:
    - joke
    - webgpt:
        val_split: 0.1
    - gpt4all:
        val_split: 0.01
    - alpaca:
        val_split: 0.025
    - code_alpaca:
        val_split: 0.05
    - minimath
    - humaneval_mbpp_codegen_qa
    - humaneval_mbpp_testgen_qa
    - grade_school_math_instructions
    - recipes
    - cmu_wiki_qa
    - oa_wiki_qa_bart_10000row
    - prosocial_dialogue:
        fraction: 0.1
    - explain_prosocial:
        fraction: 0.05
    - oig_file:
        source_url: https://huggingface.co/datasets/laion/OIG/resolve/main/unified_chip2.jsonl
        max_count: 10000
        min_length: 250
        val_split: 0.1

pythia:

pythia-2.8b-pretrain:
  dtype: fp16
  learning_rate: 6e-6
  model_name: EleutherAI/pythia-2.8b-deduped 
  deepspeed_config: configs/zero3_config_pretrain.json
  weight_decay: 0.0
  max_length: 2048
  use_flash_attention: true
  warmup_steps: 50
  gradient_checkpointing: false
  gradient_accumulation_steps: 1
  per_device_train_batch_size: 12
  per_device_eval_batch_size: 12
  num_train_epochs: 2 
  save_total_limit: 2

command: deepspeed trainer_sft.py --configs defaults pretrain pythia-2.8b-pretrain --cache_dir .cache/ --output_dir .saved_models/pythia-2.8b-pre --residual_dropout 0.0 --deepspeed