New run with DeepSpeed 0.7.7 and transformers 4.26.1, (using WarmupDecayLR, probably not optimal).

wandb run: https://wandb.ai/open-assistant/supervised-finetuning/runs/bxyaxo4v Checkpoint: 2000 steps ~48% 1st epoch

Datasets:

pretrain:
 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
 #- youtube_subs_howto100m # uses incompatible column names
 #- ubuntu_dialogue_qa # fails to load
 - oa_wiki_qa_bart_10000row
 - prosocial_dialogue:
 fraction: 0.1
 - explain_prosocial:
 fraction: 0.05

Pythia:

pythia-12b:
 dtype: fp16
 log_dir: "pythia_log_12b"
 learning_rate: 6e-6
 model_name: EleutherAI/pythia-12b-deduped
 output_dir: pythia_model_12b
 weight_decay: 0.0
 max_length: 2048
 use_flash_attention: true
 deepspeed_config: configs/zero_conf2.json
 warmup_steps: 50
 gradient_checkpointing: true
 gradient_accumulation_steps: 2
 per_device_train_batch_size: 8 
 per_device_eval_batch_size: 5
 eval_steps: 200
 save_steps: 500
 num_train_epochs: 2 
 save_total_limit: 2

zero_conf2.json:

{
 "fp16": {
 "enabled": "auto",
 "loss_scale": 0,
 "loss_scale_window": 1000,
 "initial_scale_power": 16,
 "hysteresis": 2,
 "min_loss_scale": 1
 },
 "bf16": {
 "enabled": "auto"
 },
 "optimizer": {
 "type": "AdamW",
 "params": {
 "lr": "auto",
 "betas": "auto",
 "eps": "auto",
 "weight_decay": "auto"
 }
 },
 "scheduler": {
 "type": "WarmupDecayLR",
 "params": {
 "warmup_min_lr": "auto",
 "warmup_max_lr": "auto",
 "warmup_num_steps": "auto",
 "total_num_steps": "auto"
 }
 },
 "zero_optimization": {
 "stage": 2,
 "allgather_partitions": true,
 "allgather_bucket_size": 1000000000.0,
 "overlap_comm": false,
 "reduce_scatter": true,
 "reduce_bucket_size": 1000000000.0,
 "contiguous_gradients": true
 },
 "gradient_accumulation_steps": "auto",
 "gradient_clipping": "auto",
 "steps_per_print": 2000,
 "train_batch_size": "auto",
 "train_micro_batch_size_per_gpu": "auto",
 "wall_clock_breakdown": false
}