Axolotl config to train a Mistral-7B base using QLoRA on 24GB GPU

mistral-7B-qlora.yaml

base_model: /path/to/Mistral-7B-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: /path/to/Sensei/created/dataset.jsonl
    type:
      system_prompt: "SYSTEM: {system}"
      field_system: system
      format: "USER: {instruction}\nASSISTANT: "
      no_input_format: "USER: {instruction}\nASSISTANT: "
      field_output: response
    

dataset_prepared_path: /path/to/a/folder/to/save/prepped_datasets
val_set_size: 0
output_dir: /your/trained/model/path

adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project: axolotl-training-runs
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 3
micro_batch_size: 2
num_epochs: 2
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:
eval_sample_packing: true
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
evals_per_epoch: 2
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"