llm-playground/notebooks/axolotl/config/huggyllama/qlora.yml

# Based on https://gist.github.com/fearnworks/723709806cebc67bafe1eb8138e7efbd
base_model: huggyllama/llama-7b
base_model_config: huggyllama/llama-7b
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
load_in_8bit: false
load_in_4bit: true
strict: false
push_dataset_to_hub:
datasets:
  # - path: AtlasUnified/Code-Instruct-Sets
  #   data_files:
  #     - unmasked-set-1.jsonl
  #     - unmasked-set-2.jsonl
  #     - unmasked-set-3.jsonl
  #     - unmasked-set-4.jsonl
  #   type: alpaca_code_instruct
  # - path: winglian/pygmalion-cleaned
  #   data_files:
  #     - v13_no_ai.cleaned.jsonl
  #   type: pygmalion
  #   shards: 4
  # - path: winglian/evals
  #   data_files:
  #     - hf/ARC-Challenge.jsonl
  #     - hf/ARC-Easy.jsonl
  #     - hf/riddle_sense.jsonl
  #   type: explainchoice:chat
  # - path: winglian/evals
  #   data_files:
  #     - hf/gsm8k.jsonl
  #     - custom/logic_inference_oa.jsonl
  #   type: alpaca_chat.load_qa
  # - path: winglian/evals
  #   data_files:
  #     - custom/in_context_qa.jsonl
  #   type: context_qa
  # - path: winglian/evals
  #   data_files:
  #     - custom/in_context_qa.jsonl
  #   type: context_qa.load_404
  # - path: winglian/evals
  #   data_files:
  #     - custom/jokes_explained_500up.jsonl
  #   type: sharegpt_jokes
  # - path: winglian/evals
  #   data_files:
  #     - custom/classify-self-chat.sharegpt.jsonl
  #     - custom/coding-self-chat.sharegpt.jsonl
  #     - custom/prose-gpt4.sharegpt.jsonl
  #     - custom/prose-rewrite-gpt4.sharegpt.jsonl
  #   type: sharegpt_simple
  # - path: winglian/evals
  #   data_files:
  #     - custom/guanaco-cleaned.en.jsonl
  #   type: sharegpt_simple.load_guanaco
  # - path: winglian/evals
  #   data_files:
  #     - openai/tldr.jsonl
  #   type: summarizetldr:chat
  # - path: winglian/evals
  #   data_files:
  #     - hellaswag/hellaswag.jsonl
  #   type: explainchoice:chat
  #   shards: 60
  # - path: metaeval/ScienceQA_text_only
  #   type: concisechoice:chat
  #   shards: 13
  # - path: teknium/GPTeacher-General-Instruct
  #   data_files: 
  #     - gpt4-instruct-similarity-0.6-dataset.json
  #   type: gpteacher:chat
  - path: QingyiSi/Alpaca-CoT
    data_files:
      # - chain-of-thought/formatted_cot_data/aqua_train.jsonl
      # - Chain-of-Thought/formatted_cot_data/creak_train.json
      # - Chain-of-Thought/formatted_cot_data/ecqa_train.json
      # - Chain-of-Thought/formatted_cot_data/esnli_train.json
      - Chain-of-Thought/formatted_cot_data/gsm8k_train.json
      # - Chain-of-Thought/formatted_cot_data/qasc_train.json
      # - Chain-of-Thought/formatted_cot_data/qed_train.json
      # - Chain-of-Thought/formatted_cot_data/sensemaking_train.json
      # - Chain-of-Thought/formatted_cot_data/strategyqa_train.json
      # - GPTeacher/Roleplay/formatted_roleplay-similarity_0.6-instruct-dataset.json
    type: "alpaca:chat"
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
adapter: qlora
lora_model_dir:
sequence_len: 2048
max_packed_sequence_len: 2048
lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project: huggyllama-qlora
wandb_watch:
wandb_run_id:
wandb_log_model: checkpoint
output_dir: ./qlora-out
batch_size: 8
micro_batch_size: 4
num_epochs: 3
optimizer: paged_adamw_32bit
torchdistx_path:
lr_scheduler: cosine
learning_rate: 0.0002
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: true
gradient_checkpointing: true
# stop training after this many evaluation losses have increased in a row
# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
early_stopping_patience: 3
resume_from_checkpoint:
auto_resume_from_checkpoints: true
local_rank:
logging_steps: 1
xformers_attention: false
flash_attention:
gptq_groupsize:
gptq_model_v1:
warmup_steps: 10
eval_steps: 5
save_steps: 10
debug:
deepspeed:
weight_decay: 0.000001
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"