Skip to content

Instantly share code, notes, and snippets.

@monk1337
Last active February 21, 2024 15:48
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save monk1337/b7ee08781d62e351db7fc7c6fe0645e0 to your computer and use it in GitHub Desktop.
Save monk1337/b7ee08781d62e351db7fc7c6fe0645e0 to your computer and use it in GitHub Desktop.
base_model: google/gemma-7b
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
load_in_8bit: false
load_in_4bit: true
strict: false
# huggingface repo
datasets:
- path: vicgalle/alpaca-gpt4
type: alpaca
val_set_size: 0.1
output_dir: ./out
adapter: qlora
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
sequence_len: 4000
sample_packing: false
pad_to_sequence_len: false
wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 3
micro_batch_size: 2
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_ratio: 0.5
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
#default deepspeed, can use more aggresive if needed like zero2, zero3
deepspeed: deepspeed_configs/zero1.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment