Merge pull request #186 from SinclairHudson/adding-tests

Moving metrics to metrics and adding pass/fail LLM tests

Author: benjaminye

llmtune/cli/toolkit.py

@@ -15,8 +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
-from llmtune.qa.qa_tests import QaTestRegistry
+from llmtune.qa.generics import LLMMetricSuite
+from llmtune.qa.qa_metrics import QaMetricRegistry
 from llmtune.ui.rich_ui import RichUI
 from llmtune.utils.ablation_utils import generate_permutations
 from llmtune.utils.save_utils import DirectoryHelper
@@ -91,10 +91,10 @@ def run_one_experiment(config: Config, config_path: Path) -> None:
     qa_file_path = dir_helper.save_paths.qa_file
     if not qa_file_path.exists():
         llm_metrics = config.qa.llm_metrics
-        tests = QaTestRegistry.create_tests_from_list(llm_metrics)
-        test_suite = LLMTestSuite.from_csv(results_file_path, tests)
-        test_suite.save_test_results(qa_file_path)
-        test_suite.print_test_results()
+        tests = QaMetricRegistry.create_tests_from_list(llm_metrics)
+        test_suite = LLMMetricSuite.from_csv(results_file_path, tests)
+        test_suite.save_metric_results(qa_file_path)
+        test_suite.print_metric_results()
 
 
 @app.command("run")

llmtune/qa/generics.py

@@ -1,84 +1,78 @@
 import statistics
-from abc import ABC, abstractmethod
 from pathlib import Path
 from typing import Dict, List, Union
 
 import pandas as pd
 
+from llmtune.qa.qa_metrics import LLMQaMetric
 from llmtune.ui.rich_ui import RichUI
 
 
-class LLMQaTest(ABC):
-    @property
-    @abstractmethod
-    def test_name(self) -> str:
-        pass
-
-    @abstractmethod
-    def get_metric(self, prompt: str, grount_truth: str, model_pred: str) -> Union[float, int, bool]:
-        pass
-
+class LLMMetricSuite:
+    """
+    Represents and runs a suite of metrics on a set of prompts,
+    golden responses, and model predictions.
+    """
 
-class LLMTestSuite:
     def __init__(
         self,
-        tests: List[LLMQaTest],
+        metrics: List[LLMQaMetric],
         prompts: List[str],
         ground_truths: List[str],
         model_preds: List[str],
     ) -> None:
-        self.tests = tests
+        self.metrics = metrics
         self.prompts = prompts
         self.ground_truths = ground_truths
         self.model_preds = model_preds
 
-        self._results = {}
+        self._results: Dict[str, List[Union[float, int]]] = {}
 
     @staticmethod
     def from_csv(
         file_path: str,
-        tests: List[LLMQaTest],
+        metrics: List[LLMQaMetric],
         prompt_col: str = "Prompt",
         gold_col: str = "Ground Truth",
         pred_col="Predicted",
-    ) -> "LLMTestSuite":
+    ) -> "LLMMetricSuite":
         results_df = pd.read_csv(file_path)
         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)
+        return LLMMetricSuite(metrics, prompts, ground_truths, model_preds)
 
-    def run_tests(self) -> Dict[str, List[Union[float, int, bool]]]:
-        test_results = {}
-        for test in self.tests:
-            metrics = []
+    def compute_metrics(self) -> Dict[str, List[Union[float, int]]]:
+        results = {}
+        for metric in self.metrics:
+            metric_results = []
             for prompt, ground_truth, model_pred in zip(self.prompts, self.ground_truths, self.model_preds):
-                metrics.append(test.get_metric(prompt, ground_truth, model_pred))
-            test_results[test.test_name] = metrics
+                metric_results.append(metric.get_metric(prompt, ground_truth, model_pred))
+            results[metric.metric_name] = metric_results
 
-        self._results = test_results
-        return test_results
+        self._results = results
+        return results
 
     @property
-    def test_results(self):
-        return self._results if self._results else self.run_tests()
+    def metric_results(self) -> Dict[str, List[Union[float, int]]]:
+        return self._results if self._results else self.compute_metrics()
 
-    def print_test_results(self):
-        result_dictionary = self.test_results
+    def print_metric_results(self):
+        result_dictionary = self.metric_results
         column_data = {key: list(result_dictionary[key]) for key in result_dictionary}
         mean_values = {key: statistics.mean(column_data[key]) for key in column_data}
         median_values = {key: statistics.median(column_data[key]) for key in column_data}
         stdev_values = {key: statistics.stdev(column_data[key]) for key in column_data}
         # Use the RichUI class to display the table
         RichUI.qa_display_metric_table(result_dictionary, mean_values, median_values, stdev_values)
 
-    def save_test_results(self, path: str):
+    def save_metric_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 = pd.DataFrame(self.metric_results)
         resultant_dataframe.to_csv(path, index=False)

llmtune/qa/qa_metrics.py

@@ -0,0 +1,210 @@
+from abc import ABC, abstractmethod
+from typing import List, Union
+
+import nltk
+import numpy as np
+import torch
+from langchain.evaluation import JsonValidityEvaluator
+from nltk import pos_tag
+from nltk.corpus import stopwords
+from nltk.tokenize import word_tokenize
+from rouge_score import rouge_scorer
+from transformers import DistilBertModel, DistilBertTokenizer
+
+
+json_validity_evaluator = JsonValidityEvaluator()
+
+nltk.download("stopwords")
+nltk.download("punkt")
+nltk.download("averaged_perceptron_tagger")
+
+
+class LLMQaMetric(ABC):
+    """
+    Abstract base class for a metric. A metric can be computed over a single
+    data instance, and outputs a scalar value (integer or float).
+    """
+
+    @property
+    @abstractmethod
+    def metric_name(self) -> str:
+        pass
+
+    @abstractmethod
+    def get_metric(self, prompt: str, grount_truth: str, model_pred: str) -> Union[float, int]:
+        pass
+
+
+class QaMetricRegistry:
+    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, metric_names: List[str]) -> List[LLMQaMetric]:
+        return [cls.registry[test]() for test in metric_names]
+
+
+@QaMetricRegistry.register("summary_length")
+class LengthMetric(LLMQaMetric):
+    @property
+    def metric_name(self) -> str:
+        return "summary_length"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> Union[float, int, bool]:
+        return abs(len(ground_truth) - len(model_prediction))
+
+
+@QaMetricRegistry.register("jaccard_similarity")
+class JaccardSimilarityMetric(LLMQaMetric):
+    @property
+    def metric_name(self) -> str:
+        return "jaccard_similarity"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> Union[float, int, bool]:
+        set_ground_truth = set(ground_truth.lower())
+        set_model_prediction = set(model_prediction.lower())
+
+        intersection_size = len(set_ground_truth.intersection(set_model_prediction))
+        union_size = len(set_ground_truth.union(set_model_prediction))
+
+        similarity = intersection_size / union_size if union_size != 0 else 0
+        return float(similarity)
+
+
+@QaMetricRegistry.register("dot_product")
+class DotProductSimilarityMetric(LLMQaMetric):
+    """Encodes both the ground truth and model prediction using DistilBERT, and
+    computes the dot product similarity between the two embeddings."""
+
+    def __init__(self):
+        model_name = "distilbert-base-uncased"
+        self.tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+        self.model = DistilBertModel.from_pretrained(model_name)
+
+    @property
+    def metric_name(self) -> str:
+        return "dot_product"
+
+    def _encode_sentence(self, sentence):
+        tokens = self.tokenizer(sentence, return_tensors="pt")
+        with torch.no_grad():
+            outputs = self.model(**tokens)
+        return outputs.last_hidden_state.mean(dim=1).squeeze().numpy()
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> Union[float, int, bool]:
+        embedding_ground_truth = self._encode_sentence(ground_truth)
+        embedding_model_prediction = self._encode_sentence(model_prediction)
+        dot_product_similarity = np.dot(embedding_ground_truth, embedding_model_prediction)
+        return float(dot_product_similarity)
+
+
+@QaMetricRegistry.register("rouge_score")
+class RougeScoreMetric(LLMQaMetric):
+    @property
+    def metric_name(self) -> str:
+        return "rouge_score"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> Union[float, int, bool]:
+        scorer = rouge_scorer.RougeScorer(["rouge1"], use_stemmer=True)
+        scores = scorer.score(model_prediction, ground_truth)
+        return float(scores["rouge1"].precision)
+
+
+@QaMetricRegistry.register("word_overlap")
+class WordOverlapMetric(LLMQaMetric):
+    @property
+    def metric_name(self) -> str:
+        return "word_overlap"
+
+    def _remove_stopwords(self, text: str) -> str:
+        stop_words = set(stopwords.words("english"))
+        word_tokens = word_tokenize(text)
+        filtered_text = [word for word in word_tokens if word.lower() not in stop_words]
+        return " ".join(filtered_text)
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> Union[float, int, bool]:
+        cleaned_model_prediction = self._remove_stopwords(model_prediction)
+        cleaned_ground_truth = self._remove_stopwords(ground_truth)
+
+        words_model_prediction = set(cleaned_model_prediction.split())
+        words_ground_truth = set(cleaned_ground_truth.split())
+
+        common_words = words_model_prediction.intersection(words_ground_truth)
+        overlap_percentage = (len(common_words) / len(words_ground_truth)) * 100
+        return float(overlap_percentage)
+
+
+@QaMetricRegistry.register("json_valid")
+class JSONValidityMetric(LLMQaMetric):
+    """
+    Checks to see if valid json can be parsed from the model output, according
+    to langchain_core.utils.json.parse_json_markdown
+    The JSON can be wrapped in markdown and this test will still pass
+    """
+
+    @property
+    def metric_name(self) -> str:
+        return "json_valid"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> float:
+        result = json_validity_evaluator.evaluate_strings(prediction=model_prediction)
+        binary_res = result["score"]
+        return float(binary_res)
+
+
+class PosCompositionMetric(LLMQaMetric):
+    def _get_pos_percent(self, text: str, pos_tags: List[str]) -> float:
+        words = word_tokenize(text)
+        tags = pos_tag(words)
+        pos_words = [word for word, tag in tags if tag in pos_tags]
+        total_words = len(text.split(" "))
+        return round(len(pos_words) / total_words, 2)
+
+
+@QaMetricRegistry.register("verb_percent")
+class VerbPercentMetric(PosCompositionMetric):
+    @property
+    def metric_name(self) -> str:
+        return "verb_percent"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> float:
+        return self._get_pos_percent(model_prediction, ["VB", "VBD", "VBG", "VBN", "VBP", "VBZ"])
+
+
+@QaMetricRegistry.register("adjective_percent")
+class AdjectivePercentMetric(PosCompositionMetric):
+    @property
+    def metric_name(self) -> str:
+        return "adjective_percent"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> float:
+        return self._get_pos_percent(model_prediction, ["JJ", "JJR", "JJS"])
+
+
+@QaMetricRegistry.register("noun_percent")
+class NounPercentMetric(PosCompositionMetric):
+    @property
+    def metric_name(self) -> str:
+        return "noun_percent"
+
+    def get_metric(self, prompt: str, ground_truth: str, model_prediction: str) -> float:
+        return self._get_pos_percent(model_prediction, ["NN", "NNS", "NNP", "NNPS"])
+
+
+# Instantiate tests
+# length_test = LengthMetric()
+# jaccard_similarity_test = JaccardSimilarityMetric()
+# dot_product_similarity_test = DotProductSimilarityMetric()
+# rouge_score_test = RougeScoreMetric()
+# word_overlap_test = WordOverlapMetric()
+# verb_percent_test = VerbPercentMetric()
+# adjective_percent_test = AdjectivePercentMetric()
+# noun_percent_test = NounPercentMetric()