AIM Fine-tuning Standardized SFT Image Interface

This page is the SFT image parameter reference. For the surrounding workflow, see the CLI reference, recipe schema reference, and engine concepts. engine-specific launch details are documented in the engine reference.

Environment Variables

The environment variables use a common prefix of AIMFT_ to avoid overloading variable names and to indicate which variables correspond to AIM Fine-tuning parameters. Certain other environment variables such as HF_TOKEN might be used by AIM Fine-tuning, but derive from other sources and thus do not use the prefix.

Environment Variable Value Encoding

The values of the environment variables are encoded in JSON. The values can any of the basic types representable in JSON: string, integer, float, object (i.e. dictionary mapping), array (i.e. list), boolean, null. Additionally, values may be unset.

Table of SFT Image Environment Variables

Key	Type	Default value / Required	Behaviour
Required			These variables must be set.
AIMFT_TRAIN_DATA_PATH	string	Required	Path inside the container to the training data file, in the SFT Data Format.
AIMFT_MODEL_PATH	string	Required	Path inside the container to the directory that contains the model, in the Hugging Face format.
AIM_MODEL_ID	string	Required	AIM ModelID, the canonical identifier of the model weight set.
AIMFT_OUTPUT_PATH	string	Required	Path inside the container to the directory where outputs are produced.
Optional			These additional variables may be set, but are not required. If not set, they take a default value.
AIMFT_RECIPE_ID	string|unset	unset	If set, manually override recipe selection and attempt to use the recipe corresponding to this ID.
AIMFT_METHOD	string	“SFT”	The training algorithm or method to use. In SFT images this is set to “SFT” by default.
AIMFT_USE_ADAPTER	string	False	If True, chooses an adapter recipe (typically LoRA). If False, chooses a full model recipe.
AIMFT_VALIDATION_DATA_PATH	string|null	null	Path inside the container to the validation data in the SFT Data Format. A null value indicates no separate validation data file is used and validation is not done.
AIMFT_EPOCHS	integer	1	Sets the number of times to iterate over the training data.
AIMFT_SEED	integer	42	Sets the random seed.
AIMFT_EVALUATION_INTERVAL	integer	1000	Sets the number of optimization steps between evaluations.
AIMFT_LOGGING_INTERVAL	integer	25	Sets the number of optimization steps between log outputs.
AIMFT_SAVE_INTERVAL	integer	500	Sets the number of steps between intermediate checkpoint outputs. If set to -1, no checkpoints are saved.
AIMFT_LEARNING_RATE	float	Recipe-dependent	Sets the learning rate.
AIMFT_LEARNING_RATE_WARMUP_STEPS	integer	Recipe-dependent	Sets the number of initial training steps over which learning rate is ramped up.
AIMFT_PER_ACCELERATOR_BATCH_SIZE	integer	Recipe-dependent	Sets the batch size per accelerator. Total batch size is AIMFT_PER_ACCELERATOR_BATCH_SIZE * AIMFT_GRADIENT_ACCUMULATION * The level of data parallelism. The level of data parallelism is defined by the type of distributed setup, and depends on the recipe.
AIMFT_GRADIENT_ACCUMULATION	integer	Recipe-dependent	Sets the number of batches over which gradients are accumulated before taking an optimizer step.
AIMFT_LORA_ALPHA	integer	Recipe-dependent	Sets the alpha parameter of the LoRA Adapter.
AIMFT_LORA_DROPOUT	float	Recipe-dependent; 0.0 if unset and LoRA is enabled	Sets the dropout probability of the LoRA Adapter.
AIMFT_LORA_RANK	integer	Recipe-dependent	Sets the rank parameter of the LoRA Adapter.
AIMFT_MAX_CONTEXT_LENGTH	integer	Recipe-dependent	Sets the maximum context length in tokens. Examples longer than this are filtered or truncated.