Llama-2-13b SuperCOT lora checkpoints

These are my 2nd round of Llama-2-13b SuperCOT Lora checkpoints trained using QLora on the SuperCOT Dataset with different parameters closer to the llama 1 supercot.

Architecture

Model Architecture: Llama-2-13b
Training Algorithm: QLora

Training Details

Dataset: SuperCOT Dataset
Datset type: alpaca
Training Parameters: See Here
Training Environment: Axolotl
sequence_len: 4096

Uploads/merges

Thanks to these gigachads for uploading

yml

base_model: NousResearch/Llama-2-13b-hf
base_model_config: NousResearch/Llama-2-13b-hf
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  
path: kaiokendev/SuperCOT-dataset
  type: alpaca
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./qlora-out/checkpoint-4230

adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

lora_r: 8
lora_alpha: 16
lora_dropout: 0
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project:
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 3
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0003

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

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>"

Acknowledgments

Special thanks to the creators of the datasets in SuperCOT. Additionally, thanks to Kaiokendev for curating the SuperCOT dataset. Thanks to the contributors of the Axolotl.

Stuff generated from axolotl:

library_name: peft

Training procedure

The following bitsandbytes quantization config was used during training:

quant_method: bitsandbytes
load_in_8bit: False
load_in_4bit: True
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: True
bnb_4bit_compute_dtype: bfloat16

Framework versions

PEFT 0.6.0.dev0