add open LMMs evaluation

configs/agents/api_agents.yaml

@@ -6,26 +6,8 @@ gpt-4o-2024-05-13:
             model: "gpt-4o-2024-05-13"
             max_tokens: 512
 
-gpt-3.5-turbo-0613:
+glm-4v:
-    import: "./openai-chat.yaml"
+    import: "./finetuned_agent.yaml"
     parameters:
-        name: "gpt-3.5-turbo-0613"
+        name: "glm-4v"
-        body:
+        url: "http://"
-            model: "gpt-3.5-turbo-0613"
-            max_tokens: 512
-
-text-davinci-003:
-    import: "./openai-text.yaml"
-    parameters:
-        name: "text-davinci-003"
-        body:
-            model: "text-davinci-003"
-            max_tokens: 512
-
-text-davinci-002:
-    import: "./openai-text.yaml"
-    parameters:
-        name: "text-davinci-002"
-        body:
-            model: "text-davinci-002"
-            max_tokens: 512

configs/agents/finetuned_agent.yaml

@@ -0,0 +1,9 @@
+module: "src.client.agents.HTTPAgent"
+parameters:
+  prompter:
+    name: "prompt_string"
+    args:
+      suffix: "<|assistant|>\n"
+      agent_format: "<|assistant|>\n{content}\n\n"
+      text_context_limit: 9776
+  return_format: "{response[response]}"

src/client/agents/http_agent.py

@@ -1,7 +1,7 @@
 import contextlib
 import time
 import warnings
-
+import math
 import requests
 from urllib3.exceptions import InsecureRequestWarning
 
@@ -102,26 +102,78 @@ class Prompter:
     @staticmethod
     def prompt_string(
         prefix: str = "",
-        suffix: str = "AGENT:",
+        suffix: str = "<|agent|>\n",
-        system_format: str = "SYSTEM: {content}\n\n###\n\n",
+        system_format: str = "<|system|>\n{content}\n\n",
-        user_format: str = "USER: {content}\n\n",
+        user_format: str = "<|user|>\n{content}\n\n",
-        agent_format: str = "AGENT: {content}\n\n",
+        agent_format: str = "<|agent|>\n{content}\n\n",
         prompt_key: str = "prompt",
+        image_key: str = "image",
+        text_context_limit: int = 9776
     ):
-        # todo (YG): current it assumes the content is always a string, but it can be a list for multimodal support
+        def prompter(messages: List[Dict]):
-        def prompter(messages: List[Dict[str, str]]):
+            nonlocal prefix, suffix, system_format, user_format, agent_format, prompt_key, image_key, text_context_limit
-            nonlocal prefix, suffix, system_format, user_format, agent_format, prompt_key
+            
-            prompt = prefix
+            def text_token_estimation(text: str):
-            for item in messages:
+                token_count = 0
+                text = text.replace("\n\n", "\n").replace("```", "`").replace("##", "#")
+                for char in text:
+                    if not char.isalnum() and char != " ":
+                        token_count += 1
+                words = text.split()
+                for word in words:
+                    char_count = 0
+                    for char in word:
+                        if char.isalnum():
+                            char_count += 1
+                    token_count += math.ceil(char_count / 6)
+                return token_count
+            
+            def item_to_prompt(item: Dict):
+                prompt, images = "", []
                 if item["role"] == "system":
                     prompt += system_format.format(content=item["content"])
                 elif item["role"] == "user":
-                    prompt += user_format.format(content=item["content"])
+                    if isinstance(item["content"], str):
+                        prompt += user_format.format(content=item["content"])
+                    elif isinstance(item["content"], list):
+                        text_str = ""
+                        for content in item["content"]:
+                            if content["type"] == "text":
+                                text_str += content["text"]
+                            else:
+                                images.append(content["image_url"]["url"].split("file://")[-1])
+                        prompt += user_format.format(content=text_str)
                 else:
                     prompt += agent_format.format(content=item["content"])
+                return prompt, images
+            
+            prompt = prefix
+            images = []
+            for item in messages[:5]:
+                item_prompt, item_images = item_to_prompt(item)
+                prompt += item_prompt
+                images += item_images
+            if len(messages) > 5:
+                prompt_tail, images_tail = item_to_prompt(messages[-1])
+                for index in range(len(messages)-2, 4, -2):
+                    agent_item = messages[index]
+                    agent_prompt, agent_images = item_to_prompt(agent_item)
+                    user_item = messages[index-1]
+                    user_prompt, user_images = item_to_prompt(user_item)
+                    if text_token_estimation(prompt + user_prompt + agent_prompt + prompt_tail) > text_context_limit:
+                        prompt_tail = "\n\n** Earlier trajectory has been truncated **\n\n" + prompt_tail
+                        break
+                    prompt_tail = user_prompt + agent_prompt + prompt_tail
+                    images_tail = user_images + agent_images + images_tail
+                prompt += prompt_tail
+                images += images_tail
             prompt += suffix
-            print(prompt)
+            if len(images) != 1:
-            return {prompt_key: prompt}
+                raise Exception("Only one image is supported")
+            return [
+                {prompt_key: prompt},
+                {image_key: open(images[0], "rb")}
+            ]
 
         return prompter
 
@@ -175,6 +227,7 @@ class HTTPAgent(AgentClient):
         self,
         url,
         proxies=None,
+        data=None,
         body=None,
         headers=None,
         return_format="{response}",
@@ -185,9 +238,11 @@ class HTTPAgent(AgentClient):
         self.url = url
         self.proxies = proxies or {}
         self.headers = headers or {}
+        self.data = data or {}
         self.body = body or {}
         self.return_format = return_format
         self.prompter = Prompter.get_prompter(prompter)
+        self.prompter_type = prompter.get("name", "role_content_dict")
         if not self.url:
             raise Exception("Please set 'url' parameter")
 
@@ -195,15 +250,27 @@ class HTTPAgent(AgentClient):
         return self.prompter(history)
 
     def inference(self, history: List[dict]) -> str:
-        history = replace_image_url(history, keep_path=False, throw_details=False)
+        if self.prompter_type == "role_content_dict":
+            history = replace_image_url(history, keep_path=False, throw_details=False)
+        else:
+            history = replace_image_url(history, keep_path=True, throw_details=True)
         for _ in range(5):
             try:
-                body = self.body.copy()
+                if self.prompter_type == "role_content_dict":
-                body.update(self._handle_history(history))
+                    body = self.body.copy()
-                with no_ssl_verification():
+                    body.update(self._handle_history(history))
-                    resp = requests.post(
+                    with no_ssl_verification():
-                        self.url, json=body, headers=self.headers, proxies=self.proxies, timeout=120
+                        resp = requests.post(
-                    )
+                            self.url, json=body, headers=self.headers, proxies=self.proxies, timeout=180
+                        )
+                else:
+                    messages = self._handle_history(history)
+                    data = self.data.copy()
+                    data.update(messages[0])
+                    with no_ssl_verification():
+                        resp = requests.post(
+                            self.url, data=data, files=messages[1], headers=self.headers, proxies=self.proxies, timeout=180
+                        )
                 # print(resp.status_code, resp.text)
                 if resp.status_code != 200:
                     # print(resp.text)