Tasks for generating and judging with LLMs¶
Working with Tasks¶
The Task is a special kind of Step that includes the LLM as a mandatory argument. As with a Step, it is normally used within a Pipeline but can also be used standalone.
For example, the most basic task is the TextGeneration task, which generates text based on a given instruction.
from distilabel.models import InferenceEndpointsLLM
from distilabel.steps.tasks import TextGeneration
task = TextGeneration(
name="text-generation",
llm=InferenceEndpointsLLM(
model_id="meta-llama/Meta-Llama-3-70B-Instruct",
tokenizer_id="meta-llama/Meta-Llama-3-70B-Instruct",
),
)
task.load()
next(task.process([{"instruction": "What's the capital of Spain?"}]))
# [
# {
# "instruction": "What's the capital of Spain?",
# "generation": "The capital of Spain is Madrid.",
# "distilabel_metadata": {
# "raw_output_text-generation": "The capital of Spain is Madrid.",
# "raw_input_text-generation": [
# {
# "role": "user",
# "content": "What's the capital of Spain?"
# }
# ],
# "statistics_text-generation": { # (1)
# "input_tokens": 18,
# "output_tokens": 8
# }
# },
# "model_name": "meta-llama/Meta-Llama-3.1-8B-Instruct"
# }
# ]
- The
LLMswill not only return the text but also astatistics_{STEP_NAME}field that will contain statistics related to the generation. If available, at least the input and output tokens will be returned.
Note
The Step.load() always needs to be executed when being used as a standalone. Within a pipeline, this will be done automatically during pipeline execution.
As shown above, the TextGeneration task adds a generation based on the instruction.
New in version 1.2.0
Since version 1.2.0, we provide some metadata about the LLM call through distilabel_metadata. This can be disabled by setting the add_raw_output attribute to False when creating the task.
Additionally, since version 1.4.0, the formatted input can also be included, which can be helpful when testing
custom templates (testing the pipeline using the dry_run method).
New in version 1.5.0
Since version 1.5.0 distilabel_metadata includes a new statistics field out of the box. The generation from the LLM will not only contain the text, but also statistics associated with the text if available, like the input and output tokens. This field will be generated with statistic_{STEP_NAME} to avoid collisions between different steps in the pipeline, similar to how raw_output_{STEP_NAME} works.
Task.print¶
New in version 1.4.0
New since version 1.4.0, Task.print Task.print method.
The Tasks include a handy method to show what the prompt formatted for an LLM would look like, let's see an example with UltraFeedback, but it applies to any other Task.
from distilabel.steps.tasks import UltraFeedback
from distilabel.models import InferenceEndpointsLLM
uf = UltraFeedback(
llm=InferenceEndpointsLLM(
model_id="meta-llama/Meta-Llama-3.1-70B-Instruct",
),
)
uf.load()
uf.print()
The result will be a rendered prompt, with the System prompt (if contained for the task) and the User prompt, rendered with rich (it will show exactly the same in a jupyter notebook).
In case you want to test with a custom input, you can pass an example to the tasksformat_input` method (or generate it on your own depending on the task), and pass it to the print method so that it shows your example:
Using a DummyLLM to avoid loading one
In case you don't want to load an LLM to render the template, you can create a dummy one like the ones we could use for testing.
from distilabel.models import LLM
from distilabel.models.mixins import MagpieChatTemplateMixin
class DummyLLM(AsyncLLM, MagpieChatTemplateMixin):
structured_output: Any = None
magpie_pre_query_template: str = "llama3"
def load(self) -> None:
pass
@property
def model_name(self) -> str:
return "test"
def generate(
self, input: "FormattedInput", num_generations: int = 1
) -> "GenerateOutput":
return ["output" for _ in range(num_generations)]
You can use this LLM just as any of the other ones to load your task and call print:
Note
When creating a custom task, the print method will be available by default, but it is limited to the most common scenarios for the inputs. If you test your new task and find it's not working as expected (for example, if your task contains one input consisting of a list of texts instead of a single one), you should override the _sample_input method. You can inspect the UltraFeedback source code for this.
Specifying the number of generations and grouping generations¶
All the Tasks have a num_generations attribute that allows defining the number of generations that we want to have per input. We can update the example above to generate 3 completions per input:
from distilabel.models import InferenceEndpointsLLM
from distilabel.steps.tasks import TextGeneration
task = TextGeneration(
name="text-generation",
llm=InferenceEndpointsLLM(
model_id="meta-llama/Meta-Llama-3-70B-Instruct",
tokenizer_id="meta-llama/Meta-Llama-3-70B-Instruct",
),
num_generations=3,
)
task.load()
next(task.process([{"instruction": "What's the capital of Spain?"}]))
# [
# {
# 'instruction': "What's the capital of Spain?",
# 'generation': 'The capital of Spain is Madrid.',
# 'distilabel_metadata': {'raw_output_text-generation': 'The capital of Spain is Madrid.'},
# 'model_name': 'meta-llama/Meta-Llama-3-70B-Instruct'
# },
# {
# 'instruction': "What's the capital of Spain?",
# 'generation': 'The capital of Spain is Madrid.',
# 'distilabel_metadata': {'raw_output_text-generation': 'The capital of Spain is Madrid.'},
# 'model_name': 'meta-llama/Meta-Llama-3-70B-Instruct'
# },
# {
# 'instruction': "What's the capital of Spain?",
# 'generation': 'The capital of Spain is Madrid.',
# 'distilabel_metadata': {'raw_output_text-generation': 'The capital of Spain is Madrid.'},
# 'model_name': 'meta-llama/Meta-Llama-3-70B-Instruct'
# }
# ]
In addition, we might want to group the generations in a single output row as maybe one downstream step expects a single row with multiple generations. We can achieve this by setting the group_generations attribute to True:
from distilabel.models import InferenceEndpointsLLM
from distilabel.steps.tasks import TextGeneration
task = TextGeneration(
name="text-generation",
llm=InferenceEndpointsLLM(
model_id="meta-llama/Meta-Llama-3-70B-Instruct",
tokenizer_id="meta-llama/Meta-Llama-3-70B-Instruct",
),
num_generations=3,
group_generations=True
)
task.load()
next(task.process([{"instruction": "What's the capital of Spain?"}]))
# [
# {
# 'instruction': "What's the capital of Spain?",
# 'generation': ['The capital of Spain is Madrid.', 'The capital of Spain is Madrid.', 'The capital of Spain is Madrid.'],
# 'distilabel_metadata': [
# {'raw_output_text-generation': 'The capital of Spain is Madrid.'},
# {'raw_output_text-generation': 'The capital of Spain is Madrid.'},
# {'raw_output_text-generation': 'The capital of Spain is Madrid.'}
# ],
# 'model_name': 'meta-llama/Meta-Llama-3-70B-Instruct'
# }
# ]
Defining custom Tasks¶
We can define a custom step by creating a new subclass of the Task and defining the following:
-
inputs: is a property that returns a list of strings with the names of the required input fields or a dictionary in which the keys are the names of the columns and the values are boolean indicating whether the column is required or not. -
format_input: is a method that receives a dictionary with the input data and returns aChatTypefollowing the chat-completion OpenAI message formatting. -
outputs: is a property that returns a list of strings with the names of the output fields or a dictionary in which the keys are the names of the columns and the values are boolean indicating whether the column is required or not. This property should always includemodel_nameas one of the outputs since that's automatically injected from the LLM. -
format_output: is a method that receives the output from theLLMand optionally also the input data (which may be useful to build the output in some scenarios), and returns a dictionary with the output data formatted as needed i.e. with the values for the columns inoutputs. Note that there's no need to include themodel_namein the output.
When using the Task class inheritance method for creating a custom task, we can also optionally override the Task.process method to define a more complex processing logic involving an LLM, as the default one just calls the LLM.generate method once previously formatting the input and subsequently formatting the output. For example, EvolInstruct task overrides this method to call the LLM.generate multiple times (one for each evolution).
from typing import Any, Dict, List, Union, TYPE_CHECKING
from distilabel.steps.tasks import Task
if TYPE_CHECKING:
from distilabel.steps.typing import StepColumns
from distilabel.steps.tasks.typing import ChatType
class MyCustomTask(Task):
@property
def inputs(self) -> "StepColumns":
return ["input_field"]
def format_input(self, input: Dict[str, Any]) -> "ChatType":
return [
{
"role": "user",
"content": input["input_field"],
},
]
@property
def outputs(self) -> "StepColumns":
return ["output_field", "model_name"]
def format_output(
self, output: Union[str, None], input: Dict[str, Any]
) -> Dict[str, Any]:
return {"output_field": output}
If your task just needs a system prompt, a user message template and a way to format the output given by the LLM, then you can use the @task decorator to avoid writing too much boilerplate code.
from typing import Any, Dict, Union
from distilabel.steps.tasks import task
@task(inputs=["input_field"], outputs=["output_field"])
def MyCustomTask(output: Union[str, None], input: Union[Dict[str, Any], None] = None) -> Dict[str, Any]:
"""
---
system_prompt: |
My custom system prompt
user_message_template: |
My custom user message template: {input_field}
---
"""
# Format the `LLM` output here
return {"output_field": output}
Warning
Most Tasks reuse the Task.process method to process the generations, but if a new Task defines a custom process method, like happens for example with Magpie, one hast to deal with the statistics returned by the LLM.