Common misconceptions.

airoboros/instructors/misconception.py

@@ -0,0 +1,7 @@
+from airoboros.instructors.inline_qa import generate as generate_inline
+
+
+async def generate(instructor, **kwargs):
+    """Generator for misconception training data."""
+    async for item in generate_inline(instructor, "misconception", **kwargs):
+        yield item

airoboros/instructors/prompts/misconception.txt

@@ -0,0 +1,25 @@
+I need you to help me generate some question answer pairs related to common misconceptions.
+
+Here are a few examples:
+Example 1: QUESTION: What does searing do to meat?
+ANSWER: While it is commonly believed that searing helps seal moisture into meat, in reality it causes the meat to lose some moisture. Meat is seared to brown it, to affect its color, flavor, and texture.
+
+Example 2: QUESTION: what do banana flavored candies taste like?
+ANSWER: One would think banana-flavored candy would taste like bananas, but as it turns out, banana-flavored candy was not intended to mimic the taste of a formerly popular variety of banana. The reason banana candy tastes different from bananas is that it is mainly flavored with only one of the many flavors a banana has, the ester compound isoamyl acetate.
+
+Example 3: QUESTION: What was the first feature-length animated film?
+ANSWER: El Apóstol, a lost 1917 Argentine silent film that used cutout animation, is considered the first feature-length animated film.  The film many commonly believe to be the first, however, is Walt Disney Studios' Snow White and the Seven Dwarfs, due to it being the first feature-length animated film to use cel animation.
+
+Please write {batch_size} more examples of a similar format, with "QUESTION: [question]" and "ANSWER: [answer]" with the correct answer.
+
+All output text should be in {language}, but the exact terms "QUESTION" and "ANSWER" are special tokens that must not be translated.
+
+The output format should be:
+QUESTION: first question
+ANSWER: first question's answer
+
+QUESTION: second question
+ANSWER: second question's answer
+...
+
+Do not number the question/answer pairs, just include them in the response format indicated.

airoboros/instructors/prompts/riddle.txt

@@ -19,7 +19,7 @@ ANSWER: Day and night.
 Example 6: QUESTION: Jackie has 3 brothers.  Each brother has 2 sisters.  How many sisters does Jackie have?
 ANSWER: Jackie has one sister.  If each brother has 2 sisters, assuming Jackie is a female, that would mean there are two total female siblings, of which Jackie is one, so 2 - 1 = 1.  The number of brothers is an extraneous detail.
 
-Please write 20 more examples of a similar format, with "QUESTION: [puzzle]" and "ANSWER: [answer]" with the correct answer, reasoning step-by-step for the answer when appropriate.
+Please write {batch_size} more examples of a similar format, with "QUESTION: [puzzle]" and "ANSWER: [answer]" with the correct answer, reasoning step-by-step for the answer when appropriate.
 
 All output text should be in {language}, but the exact terms "QUESTION" and "ANSWER" are special tokens that must not be translated.
 

airoboros/self_instruct.py

@@ -763,6 +763,9 @@ async def run(self):
         from airoboros.instructors.general import generate as general_generator
         from airoboros.instructors.gtkm import generate as gtkm_generator
         from airoboros.instructors.joke import generate as joke_generator
+        from airoboros.instructors.misconception import (
+            generate as misconception_generator,
+        )
         from airoboros.instructors.multiple_choice import (
             generate as multiple_choice_generator,
         )
@@ -791,6 +794,7 @@ async def run(self):
             "experience": experience_generator,
             "general": general_generator,
             "joke": joke_generator,
+            "misconception": misconception_generator,
             "multiple_choice": multiple_choice_generator,
             "plan": plan_generator,
             "orca": orca_generator,

example-config.yaml

@@ -429,3 +429,12 @@ instructors:
       temperature: 0.7
     question_count: 15
     count: 100
+
+  ##################################################################################
+  # Common misconceptions.
+  misconception:
+    api_params:
+      temperature: 0.5
+    batch_size: 25
+    min_docsearch_score: 0.02
+    count: 500

setup.py

@@ -6,7 +6,7 @@
 
 setup(
     name="airoboros",
-    version="2.1.5",
+    version="2.1.6",
     description="Updated and improved implementation of the self-instruct system.",
     long_description=long_description,
     long_description_content_type="text/markdown",