22810c97b7
* use warmup_ratio as a better default than warmup steps since it's data dependent * replace remainder of warmup_steps
65 lines
1.1 KiB
YAML
65 lines
1.1 KiB
YAML
base_model: 01-ai/Yi-34B-Chat
|
|
# optionally might have model_type or tokenizer_type
|
|
model_type: LlamaForCausalLM
|
|
tokenizer_type: LlamaTokenizer
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
|
|
load_in_8bit: false
|
|
load_in_4bit: true
|
|
sequence_len: 1024
|
|
bf16: auto
|
|
tf32: false
|
|
flash_attention: true
|
|
special_tokens:
|
|
bos_token: "<|startoftext|>"
|
|
eos_token: "<|endoftext|>"
|
|
unk_token: "<unk>"
|
|
|
|
# Data
|
|
datasets:
|
|
- path: mhenrichsen/alpaca_2k_test
|
|
type: alpaca
|
|
warmup_ratio: 0.1
|
|
|
|
# Iterations
|
|
num_epochs: 1
|
|
|
|
# Evaluation
|
|
val_set_size: 0.1
|
|
evals_per_epoch: 5
|
|
eval_sample_packing: false
|
|
eval_batch_size: 1
|
|
|
|
# LoRA
|
|
output_dir: ./outputs/qlora-out
|
|
adapter: qlora
|
|
lora_model_dir:
|
|
lora_r: 32
|
|
lora_alpha: 16
|
|
lora_dropout: 0.05
|
|
lora_target_linear: true
|
|
lora_target_modules:
|
|
|
|
# Sampling
|
|
sample_packing: false
|
|
pad_to_sequence_len: false
|
|
|
|
# Batching
|
|
gradient_accumulation_steps: 4
|
|
micro_batch_size: 1
|
|
gradient_checkpointing: true
|
|
|
|
# wandb
|
|
wandb_project:
|
|
|
|
# Optimizer
|
|
optimizer: paged_adamw_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 0.0002
|
|
|
|
# Misc
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
weight_decay: 0
|