--- title: Config Reference 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: # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink. shrink_embeddings: # Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs embeddings_skip_upcast: # Whether to load the model with randomly initialized weights. Useful for # pre-training a model from scratch or debugging purposes. random_init_weights: # (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 the base model loading from_pretrained overrides_of_model_kwargs: # use_cache: False # 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 # quantization aware training qat: activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8" weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4" and "int8" group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after # post-training quantization quantization: weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8 activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8" group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer. # 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`, or 'auto' for automatic detection. require >=ampere # Use CUDA fp16 fp16: true # Use CUDA tf32 tf32: true # require >=ampere # Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting # 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 # List[str]. Add plugins to extend the pipeline. # See `src/axolotl/integrations` for the available plugins or doc below for more details. # https://docs.axolotl.ai/docs/custom_integrations.html plugins: # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin # A list of one or more datasets to finetune the model with # See https://docs.axolotl.ai/docs/dataset_loading.html for guide on loading datasets # See https://docs.axolotl.ai/docs/dataset-formats/ for guide on dataset formats datasets: # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory - path: vicgalle/alpaca-gpt4 # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection] type: alpaca # format | format: (chat/instruct) | .load_ 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] split dataset into N pieces (use with shards_idx) shards_idx: # Optional[int] = 0 the index of sharded dataset to use preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`) name: # Optional[str] name of dataset configuration to load 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 chat template. Used only if `chat_template: jinja` or empty. chat_template_jinja: # Key containing the messages (default: "messages") field_messages: messages # Key containing the system message (default: "system") # If the system message is not present in the dataset sample, it will be loaded from the field_system property. field_system: system # Mapping of properties from the input dataset to the chat template. # (default: message_property_mappings={'role':'role', 'content':'content'}) # If a property exists in the template but not in this mapping, the system will attempt # to load it directly from the message using the property name as the key. # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role', # while 'value' is loaded and used as 'content' in the chat template. message_property_mappings: role: from content: value # ... # Optional[Dict[str, List]]. Roles mapping in the messages. # The format is {target_role: [source_roles]}. All source roles will be mapped to the target role. # The default is: roles: user: ["human", "user"] assistant: ["gpt", "assistant"] system: ["system"] tool: ["tool"] # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template. # This does not drop the default system message from chat_template if it exists. If you wish to, # we recommend using a custom jinja template with the default system message removed or # adding a system turn with empty content. drop_system_message: # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags # See example at `docs/dataset-formats/conversation.qmd` split_thinking: # IMPORTANT: The following fields determine which parts of the conversation to train on. # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train # See examples at `docs/dataset-formats/conversation.qmd` # Note: If the below 5 fields are empty, defaults to training only on the last message. # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss. roles_to_train: ["assistant"] # default # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are: # - all: train on all EOS tokens # - turn (default): train on the EOS token at the end of each trainable turn # - last: train on the last EOS token in the conversation # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`. train_on_eos: turn # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are: # - all: train on all EOT tokens # - turn: train on the EOT token at the end of each trainable turn # - last: train on the last EOT token in the conversation # If not specified, defaults to the value of train_on_eos for backward compatibility. train_on_eot: # 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). 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 # Deduplicates datasets and test_datasets with identical entries. dataset_exact_deduplication: 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', 'simpo', 'orpo', 'grpo' rl: rl_beta: # Optional[float]. The beta parameter for the RL training. # dpo dpo_use_weighting: # Optional[bool]. Whether to perform weighting. rpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper. # orpo orpo_alpha: 0.1 # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping. # kto kto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss. kto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss. # simpo cpo_alpha: 1.0 # Weight of the BC regularizer simpo_gamma: 0.5 # Target reward margin for the SimPO loss # grpo trl: use_vllm: # Optional[bool]. Whether to use VLLM for RL training. vllm_server_host: # Optional[str]. Host of the vLLM server to connect to. vllm_server_port: # Optional[int]. Port of the vLLM server to connect to. vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond. vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding. beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use max_completion_length: # Optional[int]. Maximum length of the completion for RL training. reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir. reward_weights: # Optional[list[float]]. List of reward weights for the reward functions. num_generations: # Optional[int]. Number of generations to sample. log_completions: # Optional[bool]. Whether to log completions. num_completions_to_print: # Optional[int]. Number of completions to print when log_completions is True. sync_ref_model: # Optional[bool]. Whether to sync the reference model. ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model. ref_model_sync_steps: # Optional[int]. Sync steps for the reference model. scale_rewards: # Optional[bool]. Whether to scale rewards by their standard deviation. temperature: # Optional[float]. Sampling temperature for the GRPO policy. top_p: # Optional[float]. Top-p sampling probability for the generation policy. top_k: # Optional[int]. Top-k sampling for the generation policy. min_p: # Optional[float]. Minimum probability for the generation policy. repetition_penalty: # Optional[float]. Penalty for tokens that appear in prompt and generated text. num_iterations: # Optional[int]. Number of iterations per batch (μ) for GRPO. epsilon: # Optional[float]. Epsilon value for clipping in the GRPO algorithm. epsilon_high: # Optional[float]. Upper-bound epsilon value for clipping in the GRPO algorithm. use_liger_loss: # Optional[bool]. Whether to use Liger loss for GRPO. loss_type: # Optional[str]. Loss formulation to use. Supported values: grpo, bnpo, dr_grpo. mask_truncated_completions: # Optional[bool]. Whether to exclude truncated completions from loss calculation. # reward modelling: `True` or `False` reward_model: # process reward modelling: `True` or `False` process_reward_model: # 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 # Optional[List[str]]. Custom EOT (End-of-Turn) tokens to mask/unmask during training. # These tokens mark the boundaries between conversation turns. # For example: ["/INST", "", "[/SYSTEM_PROMPT]"] # If not specified, defaults to just the model's eos_token. # This is useful for templates that use multiple delimiter tokens. eot_tokens: # - "" # - "[/INST]" # - "[/SYSTEM_PROMPT]" # 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: # Optional[str] repo_org/repo_name # 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 sample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially. # whether to concatenate samples during pretraining pretraining_sample_concatenation: curriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning # Use batch flattening for speedups when not using sample_packing batch_flattening: # 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 # List[int] | int. # The layer indices to transform, otherwise, apply to all layers # https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform peft_layers_to_transform: # Optional[bool]. Whether to use DoRA. # https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora peft_use_dora: # Optional[bool]. Whether to use RSLoRA. # https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora peft_use_rslora: # Optional[list[tuple[int, int]]]. List of layer indices to replicate. # https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora peft_layer_replication: # bool | Literal["gaussian", "eva", "olora", "pissa", "pissa_niter_[number of iters]", "corda", "loftq"] # How to initialize LoRA weights. Default to True which is MS original implementation. # https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization peft_init_lora_weights: # 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 # Apply custom LoRA autograd functions and activation function Triton kernels for # speed and memory savings # See: https://docs.axolotl.ai/docs/lora_optims.html lora_mlp_kernel: true lora_qkv_kernel: true lora_o_kernel: true # 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. # Tensorboard use_tensorboard: # Optional[bool] # Where to save the full-finetuned model to output_dir: ./completed-model # Whether to use torch.compile and which backend to use # setting to `auto` will enable torch compile when torch>=2.5.1 torch_compile: # Optional[Union[Literal["auto"], 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, integer for every N steps. float for fraction of total steps evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps eval_strategy: # Set to `"no"` to skip evaluation, `"epoch"` at end of each epoch, leave empty to infer from `eval_steps`. save_strategy: # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of each epoch, `"best"` when better result is achieved, leave empty to infer from `save_steps`. save_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps 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 save_only_model: # Save only the model weights, skipping the optimizer. Using this means you can't resume from checkpoints. # 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: # bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time. include_tokens_per_second: # Optional[bool] # whether to find batch size that fits in memory. Passed to underlying transformers Trainer auto_find_batch_size: # Optional[bool] 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 do_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`. eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"] profiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir. # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information # snapshots can be visualized @ https://pytorch.org/memory_viz 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. Available options are: true, false, "offload", "offload_disk". # 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 # Valid values are driven by the Transformers SchedulerType class, see: # https://github.com/huggingface/transformers/blob/5f4ecf2d9f867a1255131d2461d75793c0cf1db2/src/transformers/trainer_utils.py#L420 # Valid values include # - 'linear' # - 'cosine' (default) # - 'cosine_with_restarts' # - 'polynomial' # - 'constant' # - 'constant_with_warmup' # - 'inverse_sqrt' # - 'reduce_lr_on_plateau' # - 'cosine_with_min_lr' # - 'warmup_stable_decay' # Additional schedulers include: # - 'one_cycle' # - 'rex' lr_scheduler: 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/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189 # # 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_torch # - adamw_torch_fused (default) # - adamw_torch_xla # - adamw_torch_npu_fused # - adamw_apex_fused # - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1) # - adafactor # - adamw_anyprecision # - adamw_torch_4bit # - ademamix # - sgd # - adagrad # - adamw_bnb_8bit # - adamw_8bit # alias for adamw_bnb_8bit # - ademamix_8bit # - lion_8bit # - lion_32bit # - paged_adamw_32bit # - paged_adamw_8bit # - paged_ademamix_32bit # - paged_ademamix_8bit # - paged_lion_32bit # - paged_lion_8bit # - rmsprop # - rmsprop_bnb # - rmsprop_bnb_8bit # - rmsprop_bnb_32bit # - galore_adamw # - galore_adamw_8bit # - galore_adafactor # - galore_adamw_layerwise # - galore_adamw_8bit_layerwise # - galore_adafactor_layerwise # - lomo # - adalomo # - grokadamw # - schedule_free_adamw # - schedule_free_sgd # - apollo_adamw # - apollo_adamw_layerwise # # Additional custom optimizers include: # - optimi_adamw # - ao_adamw_8bit # - ao_adamw_fp8 # - came_pytorch 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_beta3: # only used for CAME Optimizer adam_epsilon: adam_epsilon2: # only used for CAME Optimizer # 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: # Optional[bool]. Whether to bettertransformers flash_optimum: # Note: Only one of the following attention patches can be used at a time. # For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`. # Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers: xformers_attention: # Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention: flash_attention: flash_attn_cross_entropy: # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only flash_attn_rms_norm: # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only flash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation flash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation # Optional[bool]. Whether to use scaled-dot-product attention # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html sdp_attention: # Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf s2_attention: # Optional[bool]. Whether to use low_cpu_mem_usage low_cpu_mem_usage: # Optional[str]. Resume from a specific checkpoint dir resume_from_checkpoint: # Optional[bool]. 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 ## Multimodal section # int | tuple[int, int] | None . Size to resize images to, width x height. # Will read from model/processor config if not set. image_size: # str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear". image_resize_algorithm: 'bilinear' ## End of multimodal section # 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: "" # eos_token: "" # unk_token: "" # pad_token: "[PAD]" # Optional[list[str]]. Add extra tokens to the tokenizer. tokens: # - "<|startoftext|>" # - "<|endoftext|>" # Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer. # Only works for tokens that are not part of the base vocab (aka are added_tokens). # Can be checked if they exist in tokenizer.json added_tokens. added_tokens_overrides: # Dict[int, str] # 128041: "<|im_start|>" # 128042: "<|im_end|>" # 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: # Sequence parallelism # Set to a divisor of the number of GPUs available to split sequences into chunks of equal size. # Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM. # E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized # subsequences, or set to 4 to split into four equal-sized subsequences. # See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details. sequence_parallel_degree: # Optional; strides across the key dimension. Larger values use more memory but should make training faster. # Must evenly divide the number of KV heads in your model. heads_k_stride: 1 # One of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to "varlen_llama3" # in the sample packing case, and "batch_ring" in the non-sample packing case. ring_attn_func: # 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: ```