Merge pull request #173 from EleutherAI/pr/scorers-contract-fixes

Fix typing in classifier

Author: SrGonao

delphi/scorers/classifier/classifier.py

@@ -21,8 +21,8 @@ def __init__(
         client: Client,
         verbose: bool,
         n_examples_shown: int,
-        log_prob: bool,
         seed: int = 42,
+        log_prob: bool = False,
         **generation_kwargs,
     ):
         """
@@ -143,7 +143,8 @@ def _parse(
         match = re.search(pattern, string)
         if match is None:
             raise ValueError("No match found in string")
-        predictions: list[bool | Literal[0, 1]] = json.loads(match.group(0))
+        raw_predictions: list[bool | Literal[0, 1]] = json.loads(match.group(0))
+        predictions = [bool(prediction) for prediction in raw_predictions]
         assert len(predictions) == self.n_examples_shown
         probabilities = (
             self._parse_logprobs(logprobs)

delphi/scorers/classifier/detection.py

@@ -39,6 +39,7 @@ def __init__(
             temperature=temperature,
             **generation_kwargs,
         )
+        self.log_prob = log_prob
 
     def prompt(self, examples: str, explanation: str) -> list[dict]:
         return detection_prompt(examples, explanation)

delphi/scorers/classifier/fuzz.py

@@ -51,7 +51,7 @@ def __init__(
             temperature=temperature,
             **generation_kwargs,
         )
-
+        self.log_prob = log_prob
         self.threshold = threshold
         self.fuzz_type = fuzz_type
 

delphi/scorers/classifier/intruder.py

@@ -170,13 +170,15 @@ def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
             active_examples = self.rng.sample(all_active_examples, num_active_examples)
 
             # highlights the active tokens with <<>> markers
-            majority_examples = []
+            formatted_examples = []
+            chosen_examples = []
             num_active_tokens = 0
             for example in active_examples:
                 text, _str_tokens = _prepare_text(
                     example, n_incorrect=0, threshold=0.3, highlighted=True
                 )
-                majority_examples.append(text)
+                formatted_examples.append(text)
+                chosen_examples.append(example)
                 num_active_tokens += (example.activations > 0).sum().item()
 
             avg_active_tokens_per_example = num_active_tokens // len(active_examples)
@@ -193,6 +195,7 @@ def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
                     threshold=0.3,
                     highlighted=True,
                 )
+
             elif self.type == "internal":
                 # randomly select a quantile to be the intruder, make sure it's not
                 # the same as the source quantile
@@ -224,18 +227,23 @@ def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
 
             # select a random index to insert the intruder sentence
             intruder_index = self.rng.randint(0, num_active_examples)
-            examples = (
-                majority_examples[:intruder_index]
+            formatted_examples = (
+                formatted_examples[:intruder_index]
                 + [intruder_sentence]
-                + majority_examples[intruder_index:]
+                + formatted_examples[intruder_index:]
+            )
+            examples = (
+                chosen_examples[:intruder_index]
+                + [intruder]
+                + chosen_examples[intruder_index:]
             )
 
             example_activations = [example.activations.tolist() for example in examples]
             example_tokens = [example.str_tokens for example in examples]
 
             batches.append(
                 IntruderSentence(
-                    examples=examples,
+                    examples=formatted_examples,
                     intruder_index=intruder_index,
                     chosen_quantile=active_quantile,
                     activations=example_activations,

delphi/scorers/classifier/sample.py

@@ -112,7 +112,7 @@ def _prepare_text(
     if n_incorrect == 0:
 
         def is_above_activation_threshold(i: int) -> bool:
-            return example.activations[i] >= abs_threshold
+            return bool((example.activations[i] >= abs_threshold).item())
 
         return _highlight(str_toks, is_above_activation_threshold), str_toks
 
@@ -137,6 +137,7 @@ def is_above_activation_threshold(i: int) -> bool:
     # The activating token is always ctx_len - ctx_len//4
     # so we always highlight this one, and if num_tokens_to_highlight > 1
     # we highlight num_tokens_to_highlight - 1 random ones
+    # TODO: This is wrong
     token_pos = len(str_toks) - len(str_toks) // 4
     if token_pos in tokens_below_threshold:
         random_indices = [token_pos]

delphi/scorers/embedding/example_embedding.py

@@ -118,10 +118,12 @@ def compute_batch_deltas(self, batch: Batch) -> tuple[float, float]:
             # Split the embeddings back into their components
             n_neg = len(batch.negative_examples)
             n_pos = len(batch.positive_examples)
-            negative_examples_embeddings = all_embeddings[:n_neg]
-            positive_examples_embeddings = all_embeddings[n_neg : n_neg + n_pos]
-            positive_query_embedding = all_embeddings[-2].unsqueeze(0)
-            negative_query_embedding = all_embeddings[-1].unsqueeze(0)
+            negative_examples_embeddings = torch.tensor(all_embeddings[:n_neg])
+            positive_examples_embeddings = torch.tensor(
+                all_embeddings[n_neg : n_neg + n_pos]
+            )
+            positive_query_embedding = torch.tensor(all_embeddings[-2]).unsqueeze(0)
+            negative_query_embedding = torch.tensor(all_embeddings[-1]).unsqueeze(0)
 
             # Compute the similarity between the query and the examples
             negative_similarities = self.model.similarity(
@@ -165,9 +167,11 @@ def _create_batches(
         # which are going to be used as "explanations"
         positive_train_examples = record.train
 
+        number_samples = min(len(positive_train_examples), len(record.not_active))
+
         # Sample from the not_active examples
         not_active_index = self.random.sample(
-            range(len(record.not_active)), len(positive_train_examples)
+            range(len(record.not_active)), number_samples
         )
         negative_train_examples = [record.not_active[i] for i in not_active_index]
 
@@ -192,6 +196,7 @@ def _create_batches(
             positive_query_str, _ = _prepare_text(
                 positive_query, n_incorrect=0, threshold=0.3, highlighted=True
             )
+
             # Prepare the negative query
             if self.method == "default":
                 # In the default method, we just sample a random negative example
@@ -205,6 +210,7 @@ def _create_batches(
                     threshold=0.3,
                     highlighted=True,
                 )
+
             elif self.method == "internal":
                 # In the internal method, we sample a negative example
                 # that has a different quantile as the positive query
@@ -216,13 +222,13 @@ def _create_batches(
                         range(len(positive_test_examples)), 1
                     )[0]
                     negative_query_temp = positive_test_examples[negative_query_idx]
-                    negative_query_quantile = negative_query.distance
+                    negative_query_quantile = negative_query_temp.quantile
 
                 negative_query = NonActivatingExample(
                     str_tokens=negative_query_temp.str_tokens,
                     tokens=negative_query_temp.tokens,
                     activations=negative_query_temp.activations,
-                    distance=negative_query_temp.quantile,
+                    distance=float(negative_query_temp.quantile),
                 )
                 # Because it is a converted activating example, it will highlight
                 # the activating tokens
@@ -234,15 +240,18 @@ def _create_batches(
             # that have the same quantile as the positive_query
             positive_examples = [
                 e
-                for e in positive_train_examples
+                for e in positive_test_examples
                 if e.quantile == positive_query.quantile
             ]
             if len(positive_examples) > 10:
-                positive_examples = self.random.sample(positive_examples, 10)
+                positive_examples = self.random.sample(positive_examples, 11)
             positive_examples_str = [
                 _prepare_text(e, n_incorrect=0, threshold=0.3, highlighted=True)[0]
                 for e in positive_examples
             ]
+            # if one example is the same as the positive query, remove it
+            if positive_query_str in positive_examples_str:
+                positive_examples_str.remove(positive_query_str)
 
             # negative examples
             if self.method == "default":
@@ -259,7 +268,7 @@ def _create_batches(
                 # that has the same quantile as the negative_query
                 negative_examples = [
                     e
-                    for e in positive_train_examples
+                    for e in positive_test_examples
                     if e.quantile == negative_query.distance
                 ]
                 if len(negative_examples) > 10:
@@ -275,7 +284,7 @@ def _create_batches(
                 positive_query=positive_query_str,
                 negative_query=negative_query_str,
                 quantile_positive_query=positive_query.quantile,
-                distance_negative_query=negative_query.distance,
+                distance_negative_query=float(negative_query.distance),
             )
             batches.append(batch)
         return batches

delphi/scorers/scorer.py

@@ -14,5 +14,5 @@ class ScorerResult(NamedTuple):
 
 class Scorer(ABC):
     @abstractmethod
-    def __call__(self, record: LatentRecord) -> ScorerResult:
+    async def __call__(self, record: LatentRecord) -> ScorerResult:
         pass

tests/test_scorers/test_classifier_contracts.py

@@ -0,0 +1,86 @@
+import pytest
+import torch
+
+from delphi.clients.client import Client, Response
+from delphi.latents import ActivatingExample, Latent, LatentRecord, NonActivatingExample
+from delphi.scorers import DetectionScorer, FuzzingScorer
+from delphi.scorers.scorer import ScorerResult
+
+
+class ConstantResponseClient(Client):
+    def __init__(self, text: str):
+        super().__init__(model="dummy")
+        self.text = text
+
+    async def generate(self, prompt, **kwargs):
+        return Response(text=self.text)
+
+
+def _activating_example() -> ActivatingExample:
+    return ActivatingExample(
+        tokens=torch.tensor([1, 2, 3], dtype=torch.int64),
+        activations=torch.tensor([0.0, 1.0, 0.0], dtype=torch.float32),
+        str_tokens=["a", "b", "c"],
+        quantile=1,
+    )
+
+
+def _non_activating_example() -> NonActivatingExample:
+    return NonActivatingExample(
+        tokens=torch.tensor([1, 2, 3], dtype=torch.int64),
+        activations=torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32),
+        str_tokens=["x", "y", "z"],
+        distance=0.0,
+    )
+
+
+def _record() -> LatentRecord:
+    return LatentRecord(
+        latent=Latent(module_name="layers.0", latent_index=0),
+        test=[_activating_example()],
+        not_active=[_non_activating_example()],
+        explanation="test explanation",
+    )
+
+
+@pytest.mark.asyncio
+async def test_detection_scorer_async_contract_returns_scorer_result():
+    scorer = DetectionScorer(
+        client=ConstantResponseClient("[1]"),
+        n_examples_shown=1,
+        verbose=False,
+    )
+
+    result = await scorer(_record())
+
+    assert isinstance(result, ScorerResult)
+    assert result.record.explanation == "test explanation"
+    assert len(result.score) > 0
+
+
+def test_detection_parse_casts_binary_ints_to_bool():
+    scorer = DetectionScorer(
+        client=ConstantResponseClient("[0, 1]"),
+        n_examples_shown=2,
+        verbose=False,
+    )
+
+    predictions, probabilities = scorer._parse("[0, 1]")
+
+    assert predictions == [False, True]
+    assert probabilities == [None, None]
+
+
+def test_fuzzing_call_sync_contract_and_log_prob_flag():
+    scorer = FuzzingScorer(
+        client=ConstantResponseClient("[1]"),
+        n_examples_shown=1,
+        verbose=False,
+        log_prob=True,
+    )
+
+    result = scorer.call_sync(_record())
+
+    assert scorer.log_prob is True
+    assert isinstance(result, ScorerResult)
+    assert len(result.score) > 0