Store the image moderation and text moderation logs

fastchat/conversation.py

@@ -524,30 +524,55 @@ def to_anthropic_vision_api_messages(self):
 
     def to_reka_api_messages(self):
         from fastchat.serve.vision.image import ImageFormat
+        from reka import ChatMessage, TypedMediaContent, TypedText
 
         ret = []
         for i, (_, msg) in enumerate(self.messages[self.offset :]):
             if i % 2 == 0:
                 if type(msg) == tuple:
                     text, images = msg
                     for image in images:
-                        if image.image_format == ImageFormat.URL:
-                            ret.append(
-                                {"type": "human", "text": text, "media_url": image.url}
-                            )
-                        elif image.image_format == ImageFormat.BYTES:
+                        if image.image_format == ImageFormat.BYTES:
                             ret.append(
-                                {
-                                    "type": "human",
-                                    "text": text,
-                                    "media_url": f"data:image/{image.filetype};base64,{image.base64_str}",
-                                }
+                                ChatMessage(
+                                    content=[
+                                        TypedText(
+                                            type="text",
+                                            text=text,
+                                        ),
+                                        TypedMediaContent(
+                                            type="image_url",
+                                            image_url=f"data:image/{image.filetype};base64,{image.base64_str}",
+                                        ),
+                                    ],
+                                    role="user",
+                                )
                             )
                 else:
-                    ret.append({"type": "human", "text": msg})
+                    ret.append(
+                        ChatMessage(
+                            content=[
+                                TypedText(
+                                    type="text",
+                                    text=msg,
+                                )
+                            ],
+                            role="user",
+                        )
+                    )
             else:
                 if msg is not None:
-                    ret.append({"type": "model", "text": msg})
+                    ret.append(
+                        ChatMessage(
+                            content=[
+                                TypedText(
+                                    type="text",
+                                    text=msg,
+                                )
+                            ],
+                            role="assistant",
+                        )
+                    )
 
         return ret
 

fastchat/serve/api_provider.py

@@ -1076,8 +1076,13 @@ def reka_api_stream_iter(
     api_key: Optional[str] = None,  # default is env var CO_API_KEY
     api_base: Optional[str] = None,
 ):
+    from reka.client import Reka
+    from reka import TypedText
+
     api_key = api_key or os.environ["REKA_API_KEY"]
 
+    client = Reka(api_key=api_key)
+
     use_search_engine = False
     if "-online" in model_name:
         model_name = model_name.replace("-online", "")
@@ -1094,34 +1099,27 @@ def reka_api_stream_iter(
 
     # Make requests for logging
     text_messages = []
-    for message in messages:
-        text_messages.append({"type": message["type"], "text": message["text"]})
+    for turn in messages:
+        for message in turn.content:
+            if isinstance(message, TypedText):
+                text_messages.append({"type": message.type, "text": message.text})
     logged_request = dict(request)
     logged_request["conversation_history"] = text_messages
 
     logger.info(f"==== request ====\n{logged_request}")
 
-    response = requests.post(
-        api_base,
-        stream=True,
-        json=request,
-        headers={
-            "X-Api-Key": api_key,
-        },
+    response = client.chat.create_stream(
+        messages=messages,
+        max_tokens=max_new_tokens,
+        top_p=top_p,
+        model=model_name,
     )
 
-    if response.status_code != 200:
-        error_message = response.text
-        logger.error(f"==== error from reka api: {error_message} ====")
-        yield {
-            "text": f"**API REQUEST ERROR** Reason: {error_message}",
-            "error_code": 1,
-        }
-        return
-
-    for line in response.iter_lines():
-        line = line.decode("utf8")
-        if not line.startswith("data: "):
-            continue
-        gen = json.loads(line[6:])
-        yield {"text": gen["text"], "error_code": 0}
+    for chunk in response:
+        try:
+            yield {"text": chunk.responses[0].chunk.content, "error_code": 0}
+        except:
+            yield {
+                "text": f"**API REQUEST ERROR** ",
+                "error_code": 1,
+            }

fastchat/serve/gradio_block_arena_anony.py

@@ -32,10 +32,10 @@
     get_ip,
     get_model_description_md,
 )
+from fastchat.serve.moderation.moderator import AzureAndOpenAIContentModerator
 from fastchat.serve.remote_logger import get_remote_logger
 from fastchat.utils import (
     build_logger,
-    moderation_filter,
 )
 
 logger = build_logger("gradio_web_server_multi", "gradio_web_server_multi.log")
@@ -173,7 +173,10 @@ def share_click(state0, state1, model_selector0, model_selector1, request: gr.Re
         )
 
 
-SAMPLING_WEIGHTS = {}
+SAMPLING_WEIGHTS = {
+    "gpt-4o-2024-05-13": 4,
+    "claude-3-5-sonnet-20240620": 4,
+}
 
 # target model sampling weights will be boosted.
 BATTLE_TARGETS = {}
@@ -201,6 +204,9 @@ def get_battle_pair(
     if len(models) == 1:
         return models[0], models[0]
 
+    if len(models) == 0:
+        raise ValueError("There are no models provided. Cannot get battle pair.")
+
     model_weights = []
     for model in models:
         weight = get_sample_weight(
@@ -289,7 +295,11 @@ def add_text(
     all_conv_text = (
         all_conv_text_left[-1000:] + all_conv_text_right[-1000:] + "\nuser: " + text
     )
-    flagged = moderation_filter(all_conv_text, model_list, do_moderation=True)
+
+    content_moderator = AzureAndOpenAIContentModerator()
+    flagged = content_moderator.text_moderation_filter(
+        all_conv_text, model_list, do_moderation=True
+    )
     if flagged:
         logger.info(f"violate moderation (anony). ip: {ip}. text: {text}")
         # overwrite the original text

fastchat/serve/gradio_block_arena_named.py

@@ -28,10 +28,10 @@
     get_ip,
     get_model_description_md,
 )
+from fastchat.serve.moderation.moderator import AzureAndOpenAIContentModerator
 from fastchat.serve.remote_logger import get_remote_logger
 from fastchat.utils import (
     build_logger,
-    moderation_filter,
 )
 
 logger = build_logger("gradio_web_server_multi", "gradio_web_server_multi.log")
@@ -174,19 +174,29 @@ def add_text(
                 no_change_btn,
             ]
             * 6
+            + [True]
         )
 
     model_list = [states[i].model_name for i in range(num_sides)]
-    all_conv_text_left = states[0].conv.get_prompt()
-    all_conv_text_right = states[1].conv.get_prompt()
-    all_conv_text = (
-        all_conv_text_left[-1000:] + all_conv_text_right[-1000:] + "\nuser: " + text
-    )
-    flagged = moderation_filter(all_conv_text, model_list)
-    if flagged:
-        logger.info(f"violate moderation (named). ip: {ip}. text: {text}")
-        # overwrite the original text
-        text = MODERATION_MSG
+    content_moderator = AzureAndOpenAIContentModerator()
+    text_flagged = content_moderator.text_moderation_filter(text, model_list)
+
+    if text_flagged:
+        logger.info(f"violate moderation. ip: {ip}. text: {text}")
+        content_moderator.write_to_json(get_ip(request))
+        for i in range(num_sides):
+            states[i].skip_next = True
+        gr.Warning(MODERATION_MSG)
+        return (
+            states
+            + [x.to_gradio_chatbot() for x in states]
+            + [""]
+            + [
+                no_change_btn,
+            ]
+            * 6
+            + [True]
+        )
 
     conv = states[0].conv
     if (len(conv.messages) - conv.offset) // 2 >= CONVERSATION_TURN_LIMIT:
@@ -201,6 +211,7 @@ def add_text(
                 no_change_btn,
             ]
             * 6
+            + [True]
         )
 
     text = text[:INPUT_CHAR_LEN_LIMIT]  # Hard cut-off
@@ -217,6 +228,7 @@ def add_text(
             disable_btn,
         ]
         * 6
+        + [False]
     )
 
 
@@ -295,7 +307,11 @@ def bot_response_multi(
             break
 
 
-def flash_buttons():
+def flash_buttons(dont_show_vote_buttons: bool = False):
+    if dont_show_vote_buttons:
+        yield [no_change_btn] * 4 + [enable_btn] * 2
+        return
+
     btn_updates = [
         [disable_btn] * 4 + [enable_btn] * 2,
         [enable_btn] * 6,
@@ -321,6 +337,7 @@ def build_side_by_side_ui_named(models):
     states = [gr.State() for _ in range(num_sides)]
     model_selectors = [None] * num_sides
     chatbots = [None] * num_sides
+    dont_show_vote_buttons = gr.State(False)
 
     notice = gr.Markdown(notice_markdown, elem_id="notice_markdown")
 
@@ -476,24 +493,24 @@ def build_side_by_side_ui_named(models):
     textbox.submit(
         add_text,
         states + model_selectors + [textbox],
-        states + chatbots + [textbox] + btn_list,
+        states + chatbots + [textbox] + btn_list + [dont_show_vote_buttons],
     ).then(
         bot_response_multi,
         states + [temperature, top_p, max_output_tokens],
         states + chatbots + btn_list,
     ).then(
-        flash_buttons, [], btn_list
+        flash_buttons, [dont_show_vote_buttons], btn_list
     )
     send_btn.click(
         add_text,
         states + model_selectors + [textbox],
-        states + chatbots + [textbox] + btn_list,
+        states + chatbots + [textbox] + btn_list + [dont_show_vote_buttons],
     ).then(
         bot_response_multi,
         states + [temperature, top_p, max_output_tokens],
         states + chatbots + btn_list,
     ).then(
-        flash_buttons, [], btn_list
+        flash_buttons, [dont_show_vote_buttons], btn_list
     )
 
     return states + model_selectors