config.qmd

---
title: Config options
description: A complete list of all configuration options.
---

```yaml
# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
# This can also be a relative path to a model on disk
base_model: ./llama-7b-hf
# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
base_model_ignore_patterns:
# If the base_model repo on hf hub doesn't include configuration .json files,
# You can set that here, or leave this empty to default to base_model
base_model_config: ./llama-7b-hf
# You can specify to choose a specific model revision from huggingface hub
revision_of_model:
# Optional tokenizer configuration path in case you want to use a different tokenizer
# than the one defined in the base model
tokenizer_config:
# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too
model_type: AutoModelForCausalLM
# Corresponding tokenizer for the model AutoTokenizer is a good choice
tokenizer_type: AutoTokenizer
# Trust remote code for untrusted source
trust_remote_code:
# use_fast option for tokenizer loading from_pretrained, default to True
tokenizer_use_fast:
# Whether to use the legacy tokenizer setting, defaults to True
tokenizer_legacy:
# Resize the model embeddings when new tokens are added to multiples of 32
# This is reported to improve training speed on some models
resize_token_embeddings_to_32x:

# (Internal use only)
# Used to identify which the model is based on
is_falcon_derived_model:
is_llama_derived_model:
is_qwen_derived_model:
# Please note that if you set this to true, `padding_side` will be set to "left" by default
is_mistral_derived_model:

# optional overrides to the base model configuration
overrides_of_model_config:
  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
  rope_scaling:
    type: # linear | dynamic
    factor: # float

# optional overrides to the bnb 4bit quantization configuration
# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig
bnb_config_kwargs:
  # These are default values
  llm_int8_has_fp16_weight: false
  bnb_4bit_quant_type: nf4
  bnb_4bit_use_double_quant: true


# Whether you are training a 4-bit GPTQ quantized model
gptq: true

# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
load_in_8bit: true
# Use bitsandbytes 4 bit
load_in_4bit:

# Use CUDA bf16
bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
# Use CUDA fp16
fp16: true
# Use CUDA tf32
tf32: true # require >=ampere

# No AMP (automatic mixed precision)
bfloat16: true # require >=ampere
float16: true

# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset
gpu_memory_limit: 20GiB
# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge
lora_on_cpu: true

# A list of one or more datasets to finetune the model with
datasets:
  # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
  - path: vicgalle/alpaca-gpt4
    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]
    type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
    ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file
    data_files: # Optional[str] path to source data files
    shards: # Optional[int] number of shards to split data into
    name: # Optional[str] name of dataset configuration to load
    train_on_split: train # Optional[str] name of dataset split to load from
    revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.
    trust_remote_code: # Optional[bool] Trust remote code for untrusted source

  # Custom user instruction prompt
  - path: repo
    type:
      # The below are defaults. only set what's needed if you use a different column name.
      system_prompt: ""
      system_format: "{system}"
      field_system: system
      field_instruction: instruction
      field_input: input
      field_output: output

      # Customizable to be single line or multi-line
      # Use {instruction}/{input} as key to be replaced
      # 'format' can include {input}
      format: |-
        User: {instruction} {input}
        Assistant:
      # 'no_input_format' cannot include {input}
      no_input_format: "{instruction} "

      # For `completion` datsets only, uses the provided field instead of `text` column
      field:

  # Using chat template
  - path: ...
    # Set type to `chat_template` to use this strategy
    type: chat_template
    # Specify the name of the chat template to use
    # The name of the chat template to use for training, following values are supported:
    # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.
    # - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py
    # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.
    # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
    chat_template: tokenizer_default
    # Custom jinja template for chat template. This will be only used if `chat_template` is set to `jinja` or empty (in which case chat_template is automatically set to `jinja`).
    chat_template_jinja:
    # The key in the data example that contains the messages. Default is "messages".
    field_messages: messages
    # The key in the message turn that contains the role. Default is "role".
    message_field_role: role
    # The key in the message turn that contains the content. Default is "content".
    message_field_content: content
    # Optional[Dict[str, List]]. Roles mapping for the messages.
    roles:
      user: ["human", "user"]
      assistant: ["gpt", "assistant", "ai"]
      system: ["system"]

    ## NOTE: Leaving the below empty will default to using the simple legacy tokenization strategy where only last message is trained on.

    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
    roles_to_train: ["gpt", "assistant"]
    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:
    # - all: train on all EOS tokens
    # - turn: train on the EOS token at the end of each trainable turn
    # - last: train on the last EOS token in the conversation
    train_on_eos: last
    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
    message_field_training: training
    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.
    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).
    # See example at `docs/dataset-formats/conversation.qmd`
    message_field_training_detail: train_detail


# If false, the datasets will not be shuffled and will keep their original order in `datasets`.
# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
shuffle_merged_datasets: true

# A list of one or more datasets to eval the model with.
# You can use either test_datasets, or val_set_size, but not both.
test_datasets:
  - path: /workspace/data/eval.jsonl
    ds_type: json
    # You need to specify a split. For "json" datasets the default split is called "train".
    split: train
    type: completion
    data_files:
      - /workspace/data/eval.jsonl

# use RL training: 'dpo', 'ipo', 'kto'
rl:
# whether to perform weighting if doing DPO training. Boolean.
dpo_use_weighting:

# The name of the chat template to use for training, following values are supported:
# - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.
# - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py
# - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.
# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
# The selected chat template will be saved to the tokenizer_config.json for easier inferencing
# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.
chat_template: tokenizer_default
# custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.
chat_template_jinja: null
# Changes the default system message
default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
# Axolotl attempts to save the dataset as an arrow after packing the data together so
# subsequent training attempts load faster, relative path
dataset_prepared_path: data/last_run_prepared
# Push prepared dataset to hub
push_dataset_to_hub: # repo path
# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
# if not set.
dataset_processes: # defaults to os.cpu_count() if not set
# Keep dataset in memory while preprocessing
# Only needed if cached dataset is taking too much storage
dataset_keep_in_memory:
# push checkpoints to hub
hub_model_id: # private repo path to push finetuned model
# how to push checkpoints to hub
# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy
hub_strategy:
# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
# Required to be true when used in combination with `push_dataset_to_hub`
hf_use_auth_token: # boolean
# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.
val_set_size: 0.04
# Num shards for whole dataset
dataset_shard_num:
# Index of shard to use for whole dataset
dataset_shard_idx:

# The maximum length of an input to train with, this should typically be less than 2048
# as most models have a token/context limit of 2048
sequence_len: 2048
# Pad inputs so each step uses constant sized buffers
# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
pad_to_sequence_len:
# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
sample_packing:
# Set to 'false' if getting errors during eval with sample_packing on.
eval_sample_packing:
# You can set these packing optimizations AFTER starting a training at least once.
# The trainer will provide recommended values for these values.
sample_packing_eff_est:
total_num_tokens:
# Increasing the following values helps with packing, but usually only slightly (<%1.)
# The number of samples packed at a time.
sample_packing_group_size: 100000
# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
sample_packing_bin_size: 200

# Passed through to transformers when loading the model when launched without accelerate
# Use `sequential` when training w/ model parallelism to limit memory
device_map:
# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.
max_memory:

# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
adapter: lora
# If you already have a lora model trained that you want to load, put that here.
# This means after training, if you want to test the model, you should set this to the value of `output_dir`.
# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.
lora_model_dir:

# LoRA hyperparameters
# For more details about the following options, see:
# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
#  - k_proj
#  - o_proj
#  - gate_proj
#  - down_proj
#  - up_proj
lora_target_linear: # If true, will target all linear modules
peft_layers_to_transform: # The layer indices to transform, otherwise, apply to all layers

# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.
# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994
lora_modules_to_save:
#  - embed_tokens
#  - lm_head

lora_fan_in_fan_out: false

# LoRA+ hyperparameters
# For more details about the following options, see:
# https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`
loraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.
loraplus_lr_embedding: #  loraplus learning rate for lora embedding layers. Default value is 1e-6.

peft:
  # Configuration options for loftq initialization for LoRA
  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization
  loftq_config:
    loftq_bits:  # typically 4 bits

# ReLoRA configuration
# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
relora_steps: # Number of steps per ReLoRA restart
relora_warmup_steps: # Number of per-restart warmup steps
relora_anneal_steps: # Number of anneal steps for each relora cycle
relora_prune_ratio: # threshold for optimizer magnitude when pruning
relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings

# wandb configuration if you're using it
# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
wandb_project: # Your wandb project name
wandb_entity: # A wandb Team name if using a Team
wandb_watch:
wandb_name: # Set the name of your wandb run
wandb_run_id: # Set the ID of your wandb run
wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training

# mlflow configuration if you're using it
mlflow_tracking_uri: # URI to mlflow
mlflow_experiment_name: # Your experiment name
mlflow_run_name: # Your run name
hf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry

# Comet configuration if you're using it
# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.
# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start
use_comet: # Enable or disable Comet integration.
comet_api_key: # API key for Comet. Recommended to set via `comet login`.
comet_workspace: # Workspace name in Comet. Defaults to the user's default workspace.
comet_project_name: # Project name in Comet. Defaults to Uncategorized.
comet_experiment_key: # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.
comet_mode: # Create a new experiment ("create") or log to an existing one ("get"). Default ("get_or_create") auto-selects based on configuration.
comet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.
comet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.

# Where to save the full-finetuned model to
output_dir: ./completed-model

# Whether to use torch.compile and which backend to use
torch_compile:  # bool
torch_compile_backend:  # Optional[str]

# Training hyperparameters

# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.
gradient_accumulation_steps: 1
# The number of samples to include in each batch. This is the number of samples sent to each GPU.
# Batch size per gpu = micro_batch_size * gradient_accumulation_steps
micro_batch_size: 2
eval_batch_size:
num_epochs: 4
warmup_steps: 100  # cannot use with warmup_ratio
warmup_ratio: 0.05  # cannot use with warmup_steps
learning_rate: 0.00003
lr_quadratic_warmup:
logging_steps:
eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps
save_strategy: # Set to `"no"` to skip checkpoint saves
save_steps: # Leave empty to save at each epoch
saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
save_total_limit: # Checkpoints saved at a time
# Maximum number of iterations to train for. It precedes num_epochs which means that
# if both are set, num_epochs will not be guaranteed.
# e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
max_steps:

eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]

loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)

# Save model as safetensors (require safetensors package)
save_safetensors:

# Whether to mask out or include the human's prompt from the training labels
train_on_inputs: false
# Group similarly sized data to minimize padding.
# May be slower to start, as it must download and sort the entire dataset.
# Note that training loss may have an oscillating pattern with this enabled.
group_by_length: false

# Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
gradient_checkpointing: false
# additional kwargs to pass to the trainer for gradient checkpointing
# gradient_checkpointing_kwargs:
#   use_reentrant: 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

# Specify a scheduler and kwargs to use with the optimizer
lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
lr_scheduler_kwargs:
cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)

# For one_cycle optim
lr_div_factor: # Learning rate div factor

# Specify optimizer
# Valid values are driven by the Transformers OptimizerNames class, see:
# https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134
#
# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of
# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used
# in the examples/ for your model and fine-tuning use case.
#
# Valid values for 'optimizer' include:
# - adamw_hf
# - adamw_torch
# - adamw_torch_fused
# - adamw_torch_xla
# - adamw_apex_fused
# - adopt_adamw (only for torch version >= 2.5.1)
# - adafactor
# - adamw_anyprecision
# - sgd
# - adagrad
# - adamw_bnb_8bit
# - lion_8bit
# - lion_32bit
# - paged_adamw_32bit
# - paged_adamw_8bit
# - paged_lion_32bit
# - paged_lion_8bit
# - galore_adamw
# - galore_adamw_8bit
# - galore_adafactor
# - galore_adamw_layerwise
# - galore_adamw_8bit_layerwise
# - galore_adafactor_layerwise
optimizer:
# Dictionary of arguments to pass to the optimizer
optim_args:
# For Galore Optimizers the following optim_args are available
# rank:  # type: int
# update_proj_gap  # type: int
# scale  # type: float
# proj_type:  # type: str, default = std

# The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm
optim_target_modules:
# - self_attn  # for llama
# - mlp

# Specify weight decay
weight_decay:
# adamw hyperparams
adam_beta1:
adam_beta2:
adam_epsilon:
# Gradient clipping max norm
max_grad_norm:

# Augmentation techniques
# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings
# currently only supported on Llama and Mistral
neftune_noise_alpha:

# Whether to bettertransformers
flash_optimum:
# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
xformers_attention:
# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
flash_attention:
flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
# Whether to use scaled-dot-product attention
# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
sdp_attention:
# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
s2_attention:
# Resume from a specific checkpoint dir
resume_from_checkpoint:
# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
# Be careful with this being turned on between different models.
auto_resume_from_checkpoints: false

# Don't mess with this, it's here for accelerate and torchrun
local_rank:

# Add or change special tokens.
# If you add tokens here, you don't need to add them to the `tokens` list.
special_tokens:
  # bos_token: "<s>"
  # eos_token: "</s>"
  # unk_token: "<unk>"
  # pad_token: "[PAD]"

# Add extra tokens.
tokens:

# FSDP
fsdp:
fsdp_config:

# Deepspeed config path. e.g., deepspeed_configs/zero3.json
deepspeed:

# Advanced DDP Arguments
ddp_timeout:
ddp_bucket_cap_mb:
ddp_broadcast_buffers:

# Path to torch distx for optim 'adamw_anyprecision'
torchdistx_path:

# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize
pretraining_dataset:

# Debug mode
debug:

# Seed
seed:

# Allow overwrite yml config using from cli
strict:
```