fixing llm QA methods + integration

Author: benjaminye

llmtune/cli/toolkit.py

@@ -15,7 +15,8 @@
 from llmtune.finetune.lora import LoRAFinetune
 from llmtune.inference.lora import LoRAInference
 from llmtune.pydantic_models.config_model import Config
-from llmtune.qa.generics import LLMTestSuite, QaTestRegistry
+from llmtune.qa.generics import LLMTestSuite
+from llmtune.qa.qa_tests import QaTestRegistry
 from llmtune.ui.rich_ui import RichUI
 from llmtune.utils.ablation_utils import generate_permutations
 from llmtune.utils.save_utils import DirectoryHelper
@@ -92,6 +93,7 @@ def run_one_experiment(config: Config, config_path: Path) -> None:
         tests = QaTestRegistry.create_tests_from_list(llm_tests)
         test_suite = LLMTestSuite.from_csv(results_file_path, tests)
         test_suite.save_test_results(qa_file_path)
+        test_suite.print_test_results()
 
 
 @app.command("run")

llmtune/pydantic_models/config_model.py

@@ -244,3 +244,4 @@ class Config(BaseModel):
     lora: LoraConfig
     training: TrainingConfig
     inference: InferenceConfig
+    qa: QaConfig

llmtune/qa/generics.py

@@ -1,3 +1,4 @@
+from pathlib import Path
 import statistics
 from abc import ABC, abstractmethod
 from typing import Dict, List, Union
@@ -18,23 +19,6 @@ def get_metric(self, prompt: str, grount_truth: str, model_pred: str) -> Union[f
         pass
 
 
-class QaTestRegistry:
-    registry = {}
-
-    @classmethod
-    def register(cls, *names):
-        def inner_wrapper(wrapped_class):
-            for name in names:
-                cls.registry[name] = wrapped_class
-            return wrapped_class
-
-        return inner_wrapper
-
-    @classmethod
-    def create_tests_from_list(cls, test_names: List[str]) -> List[LLMQaTest]:
-        return [cls.create_test(test) for test in test_names]
-
-
 class LLMTestSuite:
     def __init__(
         self,
@@ -51,11 +35,17 @@ def __init__(
         self._results = {}
 
     @staticmethod
-    def from_csv(file_path: str, tests: List[LLMQaTest]) -> "LLMTestSuite":
+    def from_csv(
+        file_path: str,
+        tests: List[LLMQaTest],
+        prompt_col: str = "Prompt",
+        gold_col: str = "Ground Truth",
+        pred_col="Predicted",
+    ) -> "LLMTestSuite":
         results_df = pd.read_csv(file_path)
-        prompts = results_df["prompt"].tolist()
-        ground_truths = results_df["ground_truth"].tolist()
-        model_preds = results_df["model_prediction"].tolist()
+        prompts = results_df[prompt_col].tolist()
+        ground_truths = results_df[gold_col].tolist()
+        model_preds = results_df[pred_col].tolist()
         return LLMTestSuite(tests, prompts, ground_truths, model_preds)
 
     def run_tests(self) -> Dict[str, List[Union[float, int, bool]]]:
@@ -84,5 +74,11 @@ def print_test_results(self):
 
     def save_test_results(self, path: str):
         # TODO: save these!
+        path = Path(path)
+        dir = path.parent
+
+        if not dir.exists():
+            dir.mkdir(parents=True, exist_ok=True)
+
         resultant_dataframe = pd.DataFrame(self.test_results)
         resultant_dataframe.to_csv(path, index=False)

llmtune/qa/qa_tests.py

@@ -9,7 +9,7 @@
 from rouge_score import rouge_scorer
 from transformers import DistilBertModel, DistilBertTokenizer
 
-from llmtune.qa.generics import LLMQaTest, QaTestRegistry
+from llmtune.qa.generics import LLMQaTest
 
 
 model_name = "distilbert-base-uncased"
@@ -21,6 +21,23 @@
 nltk.download("averaged_perceptron_tagger")
 
 
+class QaTestRegistry:
+    registry = {}
+
+    @classmethod
+    def register(cls, *names):
+        def inner_wrapper(wrapped_class):
+            for name in names:
+                cls.registry[name] = wrapped_class
+            return wrapped_class
+
+        return inner_wrapper
+
+    @classmethod
+    def create_tests_from_list(cls, test_names: List[str]) -> List[LLMQaTest]:
+        return [cls.registry[test]() for test in test_names]
+
+
 @QaTestRegistry.register("summary_length")
 class LengthTest(LLMQaTest):
     @property