base_model: openaccess-ai-collective/trophonius-15b
base_model_config: openaccess-ai-collective/trophonius-15b
load_in_8bit: false
load_in_4bit: false
gptq: false
strict: false
push_dataset_to_hub: winglian
hf_use_auth_token: true
datasets:
  - path: openaccess-ai-collective/flan-cot-zs-noopt
    data_files:
      - cot_zs_noopt_train_augmented.gpt4.jsonl
    type: alpaca_w_system
  - path: winglian/evals
    data_files:
      - hf/arc-c-v2.jsonl
      - hf/arc-e-v2.jsonl
      - hf/riddle_sense-v2.jsonl
    type: explainchoice:chat
  - path: winglian/evals
    data_files:
      - hellaswag/hellaswag-v2.jsonl
    type: explainchoice:chat
    shards: 5
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
adapter:
lora_model_dir:
sequence_len: 2048
max_packed_sequence_len: 2048
lora_r:
lora_alpha:
lora_dropout:
lora_target_modules:
lora_target_linear:
lora_fan_in_fan_out:
wandb_project: dodona-15b
wandb_watch:
wandb_run_id:
wandb_log_model:
output_dir: ./dodona-15b
gradient_accumulation_steps: 1
micro_batch_size: 2
num_epochs: 4
  # optimizer: adamw_bnb_8bit
optimizer: paged_adamw_8bit
adam_beta2: 0.95
torchdistx_path:
lr_scheduler: cosine
learning_rate: 0.00003
train_on_inputs: false
group_by_length: true
bf16: true
fp16: false
tf32: true
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:
gptq_groupsize:
gptq_model_v1:
warmup_steps: 100
eval_steps: 67
save_steps: 201
load_best_model_at_end: false
debug:
deepspeed:
weight_decay: 0.0001
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_offload_params: true
  fsdp_transformer_layer_cls_to_wrap: GPTBigCodeBlock
special_tokens:
  pad_token: "<|endoftext|>"
  bos_token: "<|endoftext|>"
  eos_token: "<|endoftext|>"
  unk_token: "<|endoftext|>"