Update model to gpt-5-mini in shiny_test_evaluation

karangattu · karangattu · commit 8d7904f12061 · 2025-08-21T13:24:32.000-07:00
Replaces the use of 'openai/gpt-5-nano-2025-08-07' with 'openai/gpt-5-mini-2025-08-07' for both the scorer and main model in the shiny_test_evaluation function.
diff --git a/tests/inspect-ai/scripts/evaluation.py b/tests/inspect-ai/scripts/evaluation.py
@@ -202,12 +202,12 @@ def shiny_test_evaluation() -> Task:
         Explanation: [Brief explanation focusing ONLY on how well the specified criteria were met]
         """,
         grade_pattern=r"GRADE:\s*([CPI])",
-        model=get_model("openai/gpt-5-nano-2025-08-07"),
+        model=get_model("openai/gpt-5-mini-2025-08-07"),
     )
 
     return Task(
         dataset=samples,
         solver=generate(),
         scorer=scorer,
-        model=get_model("openai/gpt-5-nano-2025-08-07"),
+        model=get_model("openai/gpt-5-mini-2025-08-07"),
     )

Original file line number	Diff line number	Diff line change
`@@ -202,12 +202,12 @@ def shiny_test_evaluation() -> Task:`
`202`	`202`	`Explanation: [Brief explanation focusing ONLY on how well the specified criteria were met]`
`203`	`203`	`""",`
`204`	`204`	`grade_pattern=r"GRADE:\s*([CPI])",`
`205`		`- model=get_model("openai/gpt-5-nano-2025-08-07"),`
	`205`	`+ model=get_model("openai/gpt-5-mini-2025-08-07"),`
`206`	`206`	`)`
`207`	`207`
`208`	`208`	`return Task(`
`209`	`209`	`dataset=samples,`
`210`	`210`	`solver=generate(),`
`211`	`211`	`scorer=scorer,`
`212`		`- model=get_model("openai/gpt-5-nano-2025-08-07"),`
	`212`	`+ model=get_model("openai/gpt-5-mini-2025-08-07"),`
`213`	`213`	`)`