fail gracefully if nvcf rejects input; compact zalgo prompts (#509)

* skip past prompts giving input value errors on nvcf

* reduce zalgo intensity to compact prompts

garak/buffs/base.py

@@ -73,9 +73,7 @@ def transform(
         """attempt copying is handled elsewhere. isn't that nice"""
         yield self._derive_new_attempt(attempt)
 
-    def untransform(
-        self, attempt: garak.attempt.Attempt
-    ) -> garak.attempt.Attempt:
+    def untransform(self, attempt: garak.attempt.Attempt) -> garak.attempt.Attempt:
         """Only necessary if buff requires a post buff hook"""
         pass
 

garak/buffs/low_resource_languages.py

@@ -74,4 +74,4 @@ def untransform(self, attempt: garak.attempt.Attempt) -> garak.attempt.Attempt:
             translated_output = response.text
             translated_outputs.append(translated_output)
         attempt.outputs = translated_outputs
-        return attempt
+        return attempt

garak/generators/base.py

@@ -105,7 +105,9 @@ def generate(self, prompt: str) -> List[str]:
                 for i in generation_iterator:
                     outputs.append(self._call_model(prompt))
 
-            cleaned_outputs = [o for o in outputs if o is not None]
+            cleaned_outputs = [
+                o for o in outputs if o is not None
+            ]  # "None" means no good response
             outputs = cleaned_outputs
 
             return outputs

garak/generators/nvcf.py

@@ -5,12 +5,14 @@
 
 """NVCF LLM interface"""
 
-import backoff
+import json
 import logging
 import os
-import requests
 import time
 
+import backoff
+import requests
+
 from garak import _config
 from garak.generators.base import Generator
 
@@ -97,7 +99,13 @@ def _call_model(self, prompt: str) -> str:
         if 400 <= response.status_code < 600:
             logging.warning("nvcf : returned error code %s", response.status_code)
             logging.warning("nvcf : returned error body %s", response.content)
-            response.raise_for_status()
+            if response.status_code == 500 and json.loads(response.content)[
+                "detail"
+            ].startswith("Input value error"):
+                logging.warning("nvcf : skipping this prompt")
+                return None
+            else:
+                response.raise_for_status()
 
         else:
             response_body = response.json()

garak/probes/encoding.py

@@ -552,7 +552,9 @@ class InjectZalgo(BaseEncodingProbe, Probe):
     def zalgo(text: bytes) -> bytes:
         from zalgolib.zalgolib import enzalgofy
 
-        zalged = enzalgofy(text=text.decode(), intensity=50)
+        zalged = enzalgofy(
+            text=text.decode(), intensity=20
+        )  # default value of 50 explodes prompt length
 
         return bytes(zalged, "utf-8")