Exemplar Selection (in context learning) API #1055
Conversation
sonichi
left a comment
sonichi
left a comment
There was a problem hiding this comment.
Could you add an example of user's code, in test or documentation?
| key + ": " + str(context[key]) for key in key_order[:-1] | ||
| ] | ||
| ) | ||
| few_shot_prompt += "\n" + key_order[-1] + ": " + "\n" |
There was a problem hiding this comment.
maybe use format string here?
There was a problem hiding this comment.
Are there any benefits to use format string?
| } | ||
|
|
||
| @classmethod | ||
| def get_few_shot_template(cls, train_data, method=None, few_shot_template=None, method_params=None, template_params=None): |
There was a problem hiding this comment.
Instead of passing selection method into exemplar_selector, maybe it would be easier to implement exemplar_selector with different methods?
BTW I'm not too familiar with pythonic-way of abstract class so I'm just make up pattern here.
abstract class ExemplarSelector:
def construct_template(context) -> str;
pass
# random_exemplar_selector
class RandomExemplarSelector(ExemplarSelector):
def __init__(self, train_data, k):
# initialize
def construct_template(context):
# select `k` examples from train_data and render template. There was a problem hiding this comment.
I add the abstract class to exemplar_selector.py
test/autogen/oai/test_selection.py
Outdated
|
|
||
| def test_case_existing_method_default_template(self): | ||
| # Most cases should use the default template and existing methods | ||
| prompt_fn = oai.ExemplarSelector.get_few_shot_template(self.exemplars, method="random", |
There was a problem hiding this comment.
Should ExemplarSelector be under oai? It doesn't look right. What about autogen/icl?
There was a problem hiding this comment.
moved to autogen/icl
| def test_case_existing_method_default_template(self): | ||
| # Most cases should use the default template and existing methods | ||
| prompt_fn = oai.ExemplarSelector.get_few_shot_template(self.exemplars, method="random", | ||
| method_params={"k": 3}, template_params={"key_order": self.key_order}) |
There was a problem hiding this comment.
In general, the field to predict in the context is not necessarily the last key. Also, the number of such fields is not necessarily 1. Make sure the prompt_fn does not depend on such assumptions.
| import unittest | ||
| import datasets | ||
| from flaml import oai | ||
| from flaml import icl |
There was a problem hiding this comment.
| from flaml import icl | |
| from flaml.autogen.icl import ExemplarSelector |
|
Close PRs for autogen. |
4 participants
Why are these changes needed?
Produced for in-context learning user cases.
For example:
Related issue number
Automate example selection in LLM few-shot learning microsoft/autogen#448
Checks