📄 Llama3 Instruct All Demos In One Turn¶

formats.llama3_instruct_all_demos_in_one_turn

SystemFormat(
    demo_format="{source}

{target_prefix}{target}

",
    model_input_format="<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}{instruction}<|eot_id|><|start_header_id|>user<|end_header_id|>

{demos}{source}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{target_prefix}",
)
[source]

Explanation about SystemFormat¶

Generates the whole input to the model, from constant strings that are given as args, and from values found in specified fields of the instance.

Important: formats can use '\N' notations that means new-line if no new-line before and no empty string before.

SystemFormat expects the input instance to contain: 1. A field named “system_prompt” whose value is a string (potentially empty) that delivers a task-independent opening text. 2. A field named “source” whose value is a string verbalizing the original values in the instance (as read from the source dataset), in the context of the underlying task. 3. A field named “instruction” that contains a (non-None) string. 4. A field named with the value in arg 'demos_field', containing a list of dicts, each dict with fields “source” and “target”, representing a single demo. 5. A field named “target_prefix” that contains a string to prefix the target in each demo, and to end the whole generated prompt

SystemFormat formats the above fields into a single string to be input to the model. This string overwrites field “source” of the instance. Formatting is driven by two args: 'demo_format' and 'model_input_format'. SystemFormat also pops fields “system_prompt”, “instruction”, “target_prefix”, and the field containing the demos out from the input instance.
Args:
demos_field (str): the name of the field that contains the demos, being a list of dicts, each with “source” and “target” keys demo_format (str): formatting string for a single demo, combining fields “source” and “target” model_input_format (str): overall product format, combining instruction and source (as read from fields “instruction” and “source” of the input instance), together with demos (as formatted into one string) format_args (Dict[str,str]): additional format args to be used when formatting the different format strings

Example:
when input instance:
{
    "source": "1+1",
    "target": "2",
    "instruction": "Solve the math exercises.",
    "demos": [{"source": "1+2", "target": "3"}, {"source": "4-2", "target": "2"}]
}
is processed by
system_format = SystemFormat(
    demos_field=constants.demos_field,
    demo_format="Input: {source}\nOutput: {target}\n\n",
    model_input_format="Instruction: {instruction}\n\n{demos}Input: {source}\nOutput: ",
)
the resulting instance is:
{
    "target": "2",
    "source": "Instruction: Solve the math exercises.\n\nInput: 1+2\nOutput: 3\n\nInput: 4-2\nOutput: 2\n\nInput: 1+1\nOutput: ",
}