moved evaluation to separate package

Aleph-Alpha · Nov 24, 2023 · 9736c8d · 9736c8d
1 parent 8a7ed05
commit 9736c8d
Show file tree

Hide file tree

Showing 9 changed files with 22 additions and 18 deletions.
diff --git a/src/intelligence_layer/core/__init__.py b/src/intelligence_layer/core/__init__.py
@@ -14,18 +14,18 @@
     Language,
 )
 from .echo import EchoInput, EchoOutput, EchoTask
-from .evaluator import Dataset as Dataset
-from .evaluator import EvaluationException as EvaluationException
-from .evaluator import EvaluationRepository as EvaluationRepository
-from .evaluator import EvaluationRunOverview as EvaluationRunOverview
-from .evaluator import Evaluator as Evaluator
-from .evaluator import Example as Example
-from .evaluator import ExampleResult as ExampleResult
-from .evaluator import InMemoryEvaluationRepository as InMemoryEvaluationRepository
-from .evaluator import LogTrace as LogTrace
-from .evaluator import SequenceDataset as SequenceDataset
-from .evaluator import SpanTrace as SpanTrace
-from .evaluator import TaskSpanTrace as TaskSpanTrace
+from .evaluation.evaluator import Dataset as Dataset
+from .evaluation.evaluator import EvaluationException as EvaluationException
+from .evaluation.evaluator import EvaluationRepository as EvaluationRepository
+from .evaluation.evaluator import EvaluationRunOverview as EvaluationRunOverview
+from .evaluation.evaluator import Evaluator as Evaluator
+from .evaluation.evaluator import Example as Example
+from .evaluation.evaluator import ExampleResult as ExampleResult
+from .evaluation.evaluator import InMemoryEvaluationRepository as InMemoryEvaluationRepository
+from .evaluation.evaluator import LogTrace as LogTrace
+from .evaluation.evaluator import SequenceDataset as SequenceDataset
+from .evaluation.evaluator import SpanTrace as SpanTrace
+from .evaluation.evaluator import TaskSpanTrace as TaskSpanTrace
 from .explain import Explain, ExplainInput, ExplainOutput
 from .graders import BleuGrader, RougeScores
 from .prompt_template import (

diff --git a/src/intelligence_layer/core/evaluation/__init__.py b/src/intelligence_layer/core/evaluation/__init__.py
diff --git a/src/intelligence_layer/core/evaluator.py → ...igence_layer/core/evaluation/evaluator.py b/src/intelligence_layer/core/evaluator.py → ...igence_layer/core/evaluation/evaluator.py
diff --git a/src/intelligence_layer/use_cases/classify/classify.py b/src/intelligence_layer/use_cases/classify/classify.py
@@ -3,7 +3,7 @@
 from pydantic import BaseModel
 
 from intelligence_layer.core.chunk import Chunk
-from intelligence_layer.core.evaluator import EvaluationRepository, Evaluator
+from intelligence_layer.core import EvaluationRepository, Evaluator
 from intelligence_layer.core.task import Task
 
 Probability = NewType("Probability", float)

diff --git a/src/intelligence_layer/use_cases/summarize/summarize.py b/src/intelligence_layer/use_cases/summarize/summarize.py
@@ -5,7 +5,7 @@
 
 from intelligence_layer.core.chunk import Chunk
 from intelligence_layer.core.detect_language import Language
-from intelligence_layer.core.evaluator import EvaluationRepository, Evaluator
+from intelligence_layer.core import EvaluationRepository, Evaluator
 from intelligence_layer.core.graders import BleuGrader, RougeGrader
 from intelligence_layer.core.task import Task
 

diff --git a/tests/core/test_evaluator.py b/tests/core/test_evaluator.py
@@ -14,7 +14,8 @@
     TaskSpanTrace,
     Tracer,
 )
-from intelligence_layer.core.evaluator import LogTrace, SpanTrace, _to_trace_entry
+from intelligence_layer.core import LogTrace, SpanTrace
+from intelligence_layer.core.evaluation.evaluator import _to_trace_entry 
 from intelligence_layer.core.task import Task
 from intelligence_layer.core.tracer import InMemorySpan, InMemoryTaskSpan, LogEntry
 
@@ -150,6 +151,9 @@ def test_evaluate_dataset_stores_aggregated_results(
     assert evaluation_run_overview == loaded_evaluation_run_overview
 
 
+
+
+
 def test_to_trace_entry() -> None:
     now = datetime.utcnow()
     entry = _to_trace_entry(

diff --git a/tests/use_cases/classify/test_embedding_based_classify.py b/tests/use_cases/classify/test_embedding_based_classify.py
@@ -11,7 +11,7 @@
     QdrantInMemoryRetriever,
 )
 from intelligence_layer.core.chunk import Chunk
-from intelligence_layer.core.evaluator import InMemoryEvaluationRepository
+from intelligence_layer.core import InMemoryEvaluationRepository
 from intelligence_layer.core.task import Task
 from intelligence_layer.core.tracer import NoOpTracer
 from intelligence_layer.use_cases.classify.classify import (

diff --git a/tests/use_cases/classify/test_single_label_classify.py b/tests/use_cases/classify/test_single_label_classify.py
@@ -6,7 +6,7 @@
     AlephAlphaClientProtocol,
 )
 from intelligence_layer.core.chunk import Chunk
-from intelligence_layer.core.evaluator import (
+from intelligence_layer.core import (
     Example,
     InMemoryEvaluationRepository,
     SequenceDataset,

diff --git a/tests/use_cases/summarize/test_summarize.py b/tests/use_cases/summarize/test_summarize.py
@@ -2,7 +2,7 @@
 
 from intelligence_layer.core.chunk import Chunk
 from intelligence_layer.core.detect_language import Language
-from intelligence_layer.core.evaluator import (
+from intelligence_layer.core import (
     Example,
     InMemoryEvaluationRepository,
     SequenceDataset,