Add Generic Instruction

prompting/task_registry.py

@@ -1,5 +1,5 @@
-from .tasks import Task, MockTask, SummarizationTask, QuestionAnsweringTask, DebuggingTask, MathTask, DateQuestionAnsweringTask
-from .tools import MockDataset, WikiDataset, HFCodingDataset, StackOverflowDataset, MathDataset, WikiDateDataset
+from .tasks import Task, MockTask, SummarizationTask, QuestionAnsweringTask, DebuggingTask, MathTask, DateQuestionAnsweringTask, GenericInstructionTask
+from .tools import MockDataset, WikiDataset, HFCodingDataset, StackOverflowDataset, MathDataset, WikiDateDataset, GenericInstructionDataset
 
 # TODO: Expand this to include extra information beyond just the task and dataset names
 mock_task, mock_dataset = MockTask.name, [MockDataset.name]
@@ -8,6 +8,7 @@
 debugging_task, debugging_dataset = DebuggingTask.name, [HFCodingDataset.name]
 math_task, math_dataset = MathTask.name, [MathDataset.name]
 date_qa_task, date_qa_dataset = DateQuestionAnsweringTask.name, [WikiDateDataset.name]
+generic_instruction_task, generic_instruction_dataset = GenericInstructionTask.name, [GenericInstructionDataset.name]
 
 TASK_REGISTRY = {
     mock_task: mock_dataset,
@@ -16,4 +17,5 @@
     debugging_task: debugging_dataset,
     math_task: math_dataset,
     date_qa_task: date_qa_dataset,
+    generic_instruction_task: generic_instruction_dataset
 }

prompting/tasks/__init__.py

@@ -14,6 +14,6 @@
     DateQuestionAnsweringTask.name: DateQuestionAnsweringTask,
     SummarizationTask.name: SummarizationTask,
     DebuggingTask.name: DebuggingTask,
-    #GenericInstructionTask.name: GenericInstructionTask,
+    GenericInstructionTask.name: GenericInstructionTask,
     MathTask.name: MathTask,
 }

prompting/tools/__init__.py

@@ -7,16 +7,18 @@
     StackOverflowDataset,
     WikiDateDataset,
     MathDataset,
+    GenericInstructionDataset,
 )
 from .selector import Selector
 
 DATASETS = {
-    "mock": MockDataset,
-    "hf_coding": HFCodingDataset,
-    "wiki": WikiDataset,
-    #"stack_overflow": StackOverflowDataset,
-    "wiki_date": WikiDateDataset,
-    "math": MathDataset,
+    MockDataset.name: MockDataset,
+    HFCodingDataset.name: HFCodingDataset,
+    WikiDataset.name: WikiDataset,
+    #StackOverflowDataset.name: StackOverflowDataset,
+    MathDataset.name: MathDataset,
+    WikiDateDataset.name: WikiDateDataset,
+    GenericInstructionDataset.name: GenericInstructionDataset, 
 }
 
 

prompting/tools/datasets/__init__.py

@@ -4,3 +4,4 @@
 from .math import MathDataset
 from .mock import MockDataset
 from .wiki import WikiDataset, WikiDateDataset
+from .generic_instruction import GenericInstructionDataset

prompting/tools/datasets/base.py

@@ -17,6 +17,8 @@
 # DEALINGS IN THE SOFTWARE.
 
 import time
+import random
+import functools
 from abc import ABC, abstractmethod
 from typing import Dict
 import bittensor as bt
@@ -28,7 +30,7 @@
 
 class Dataset(ABC):
     """Base class for datasets."""
-    name = "dataset"
+
     max_tries: int = 10
 
     @abstractmethod
@@ -74,11 +76,51 @@ def next(
                     f"Could not find any samples which meet {self.__class__.__name__} requirements after {tries} tries."
                 )
 
+        info["source"] = (self.__class__.__name__,)
         info["stats"] = {
-            "creator": self.__class__.__name__,
             "fetch_time": time.time() - t0,
             "num_tries": tries,
             "fetch_method": method,
             "next_kwargs": kwargs,
         }
         return Context(**info)
+
+
+class TemplateDataset(Dataset):
+    """Base class for datasets based on a template."""
+
+    @property
+    def size(self):
+        return functools.reduce(
+            lambda x, y: x * y, [len(v) for v in self.params.values()], 1
+        )
+
+    def __repr__(self):
+        return f"{self.__class__.__name__} with template: {self.query_template!r} and {self.size} possible phrases"
+
+    def get(self, params: dict):
+        content = self.query_template.format(**params)
+        keys, values = list(zip(*params.items()))
+
+        return {
+            "title": params.get(
+                "title", keys[0]
+            ),  # Use the first key as the title if no field called title is present
+            "topic": params.get("topic", keys[min(1, len(keys) - 1)]),  # Same for topic
+            "subtopic": params.get(
+                "subtopic", keys[min(2, len(keys) - 2)]
+            ),  # Same for subtopic
+            "content": content,  # content
+            "internal_links": values,  # internal links
+            "external_links": values,  # external links
+            "tags": values,  # tags
+            "extra": {},
+        }
+
+    def random(self, selector: Selector = None):
+        selected = {k: selector(v) for k, v in self.params.items()}
+        return self.get(selected)
+
+    def search(self, params: dict, selector: Selector = None):
+        selected = {k: params.get(k, selector(v)) for k, v in self.params.items()}
+        return self.get(selected)

prompting/tools/datasets/generic_instruction.py

@@ -1,9 +1,9 @@
 from .base import TemplateDataset
 
 
-class GenericQuestionDataset(TemplateDataset):
+class GenericInstructionDataset(TemplateDataset):
     "Generic question dataset, which creates LLM prompts for asking questions."
-
+    name = "generic_instruction"
     query_template = (
         "Ask a {style} question about a {theme} {subtopic} related to {topic}"
     )