metadata
datasets:
- mhenrichsen/alpaca_2k_test
pipeline_tag: text2text-generation
Small qlora finetune using Axolotl. Locally tested using wikitext perplexity test and had a small improvement over the base Llama v2 7B base model.
Axolotl config used:
base_model: NousResearch/Llama-2-7b-hf
base_model_config: NousResearch/Llama-2-7b-hf
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
push_dataset_to_hub:
hub_model_id:
load_in_8bit: false
load_in_4bit: true
strict: false
datasets:
- path: mhenrichsen/alpaca_2k_test
type: alpaca
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./checkpoints/llama-2-qlora
adapter: qlora
lora_model_dir:
sequence_len: 4096
max_packed_sequence_len: 4096
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project:
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002
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: true
flash_attention:
warmup_steps: 10
eval_steps: 20
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
bos_token: "<s>"
eos_token: "</s>"
unk_token: "<unk>"
And then merged with Axolotl via:
accelerate launch scripts/finetune.py configs/your_config.yml --merge_lora --lora_model_dir="./completed-model" --load_in_8bit=False --load_in_4bit=False