Saving results (hash always diffrent) = skipping doesnt work

test_small_llms.py

@@ -19,9 +19,9 @@ def main():
         # "tinyllama:1.1b",
     ]
     seeds = [
-        # 2,
-        222,
-        22222,
+        2,
+        # 222,
+        # 22222,
         # 2222222
     ]
     tests = [
@@ -34,9 +34,13 @@ def main():
     	    },
             validator=system_human_answer_match,
     	    validation_input={
-        	    "criteria": """- in Mandarin Chinese
-                            - factually correct
-                            - just one single conversational answer, without any AI fragments (A/B versions, "end of message" parts, unfitting discalimers or notes)""",
+        	    "criteria": """- in Mandarin Chinese from front to finnish
+- factually correct
+- about healthy vegetables
+- just one single conversational answer, without any AI fragments (A/B versions, "end of message" parts, unfitting discalimers or notes)
+
+Again, the message has to be entirely in Manadarin Chineese.
+That means If the answer is not in Chinese the answer is NOT correct! Only if the message in in Chinese rate as correct""",
     	    }
         ),
         Test(
@@ -53,6 +57,7 @@ def main():
             validator=regex_match_any,
     	    validation_input={
     	        "patterns": ["33549659245", "33,549,659,245", "33.549.659.245"]
+    	        # "patterns": ["3[,\. ]?3[,\. ]?5[,\. ]?4[,\. ]?9[,\. ]?6[,\. ]?5[,\. ]?9[,\. ]?2[,\. ]?4[,\. ]?5"] # Would accept 3.354.965.9245
     	    }
         ),
     ]
@@ -65,7 +70,14 @@ def main():
     )
 
     print()
-    for result in results: print(f"\n\033[0;36mtest_name:\033[0m {result['test'].name}\n\033[0;36mmodel:\033[0m {result['model']}\n\033[0;36mseed:\033[0m {result['seed']}\n\033[0;36mvalidation_result:\033[0m {result['validation']}\n\033[0;36manswer: ⏎\033[0m\n{result['result']}")
+    for hash_key in results:
+        result = results[hash_key]
+        print(f"""
+\033[0;36mtest_name:\033[0m {result['test_name']}
+\033[0;36mmodel:\033[0m {result['model']}
+\033[0;36mseed:\033[0m {result['seed']}
+\033[0;36mvalidation_result:\033[0m {result['validation']}
+\033[0;36manswer: »\033[0m{result['answer']}\033[0;36m«\033[0m""")
 
 if __name__ == "__main__":
     main()