Added support for GPT-4o

Author: sakher-sawan

evals/elsuite/track_the_stat/scripts/make_plots.py

@@ -26,12 +26,14 @@ def zero_if_none(input_num):
     "llama-2-70b-chat",
     "random_baseline",
     "human_baseline",
+    "gpt-4o"
 ]
 # separate list for OAI models for token counting, not supported in others.
 OAI_MODELS = [
     "gpt-4-0125-preview",
     "gpt-3.5-turbo-0125",
     "gpt-4-base",
+    "gpt-4o"
 ]
 
 STAT_TO_LABEL = {
@@ -50,6 +52,8 @@ def get_model(spec):
     # this is hilariously ugly but it works for now (sorry)
     if "gpt-4-turbo-preview" in spec["completion_fns"][0]:
         return "gpt-4-0125-preview"
+    if "gpt-4o" in spec["completion_fns"][0]:
+        return "gpt-4o"
     elif "gpt-3.5-turbo" in spec["completion_fns"][0]:
         return "gpt-3.5-turbo-0125"
     elif "gpt-4-base" in spec["completion_fns"][0]:

evals/registry.py

@@ -42,6 +42,7 @@ def n_ctx_from_model_name(model_name: str) -> Optional[int]:
         ("gpt-3.5-turbo-", 4096),
         ("gpt-4-32k-", 32768),
         ("gpt-4-", 8192),
+        ("gpt-4o", 128_000),
     ]
     MODEL_NAME_TO_N_CTX: dict[str, int] = {
         "ada": 2048,
@@ -65,6 +66,7 @@ def n_ctx_from_model_name(model_name: str) -> Optional[int]:
         "gpt-4-1106-preview": 128_000,
         "gpt-4-turbo-preview": 128_000,
         "gpt-4-0125-preview": 128_000,
+        "gpt-4o": 128_000
     }
 
     # first, look for an exact match
@@ -84,7 +86,7 @@ def is_chat_model(model_name: str) -> bool:
     if model_name in {"gpt-4-base"} or model_name.startswith("gpt-3.5-turbo-instruct"):
         return False
 
-    CHAT_MODEL_NAMES = {"gpt-3.5-turbo", "gpt-3.5-turbo-16k", "gpt-4", "gpt-4-32k"}
+    CHAT_MODEL_NAMES = {"gpt-3.5-turbo", "gpt-3.5-turbo-16k", "gpt-4", "gpt-4-32k", "gpt-4o"}
 
     if model_name in CHAT_MODEL_NAMES:
         return True