WIP large refactor to make finetune script a little more manageable (#3)

configs/gpt_neox_20b.yml

@@ -0,0 +1,39 @@
+base_model: EleutherAI/gpt-neox-20b
+base_model_ignore_patterns: pytorch*  # prefer safetensors
+model_type: GPTNeoXForCausalLM
+tokenizer_type: AutoTokenizer
+load_in_8bit: true
+datasets:
+  - path: nomic-ai/gpt4all-j-prompt-generations
+    type: alpaca
+    shards: 4
+    shards_index: 0
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+adapter: lora
+lora_model_dir:
+sequence_len: 2048
+max_packed_sequence_len: 2048
+lora_r: 8
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - query_key_value
+lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
+wandb_project: gpt4all-neox-20b
+wandb_watch:
+wandb_run_id:
+wandb_log_model: checkpoint
+output_dir: ./gpt4all-neox-20b
+batch_size: 48
+micro_batch_size: 4
+num_epochs: 5
+learning_rate: 0.00003
+lr_scheduler: one_cycle
+train_on_inputs: false
+group_by_length: false
+bf16: True
+tf32: True
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank: