Update space

apis/chat_api.py

@@ -55,8 +55,8 @@ class ChatAPIApp:
                     "owned_by": "NousResearch",
                 },
                 {
-                    "id": "gpt-neo-2.7B",
-                    "description": "[EleutherAI/gpt-neo-2.7B]: https://huggingface.co/EleutherAI/gpt-neo-2.7B",
+                    "id": "zephyr-7b-beta",
+                    "description": "[HuggingFaceH4/zephyr-7b-beta]: https://huggingface.co/HuggingFaceH4/zephyr-7b-beta",
                     "object": "model",
                     "created": 1700000000,
                     "owned_by": "TheBloke",

messagers/message_composer.py

@@ -8,7 +8,7 @@ class MessageComposer:
     AVALAIBLE_MODELS = [
         "mixtral-8x7b",
         "mistral-7b",
-        "gpt-neo-2.7B",
+        "zephyr-7b-beta",
         "nous-mixtral-8x7b",
     ]
 
@@ -95,27 +95,21 @@ class MessageComposer:
             self.merged_str_list.append("<|im_start|>assistant")
             self.merged_str = "\n".join(self.merged_str_list)
         # https://huggingface.co/openchat/openchat-3.5-0106
-        elif self.model in ["gpt-neo-2.7B"]:
+        elif self.model in ["zephyr-7b-beta"]:
             self.messages = self.concat_messages_by_role(messages)
-            self.merged_str_list = []
-            self.end_of_turn = "<|end_of_turn|>"
+            self.cached_str = ""
             for message in self.messages:
                 role = message["role"]
                 content = message["content"]
                 if role in self.inst_roles:
-                    self.merged_str_list.append(
-                        f"GPT4 Correct User:\n{content}{self.end_of_turn}"
-                    )
+                    self.cached_str = f"[INST] {content} [/INST]"
                 elif role in self.answer_roles:
-                    self.merged_str_list.append(
-                        f"GPT4 Correct Assistant:\n{content}{self.end_of_turn}"
-                    )
+                    self.merged_str += f"<s> {self.cached_str} {content} </s>\n"
+                    self.cached_str = ""
                 else:
-                    self.merged_str_list.append(
-                        f"GPT4 Correct User: {content}{self.end_of_turn}"
-                    )
-            self.merged_str_list.append(f"GPT4 Correct Assistant:\n")
-            self.merged_str = "\n".join(self.merged_str_list)
+                    self.cached_str = f"[INST] {content} [/INST]"
+            if self.cached_str:
+                self.merged_str += f"{self.cached_str}"
         else:
             self.merged_str = "\n".join(
                 [
@@ -191,18 +185,19 @@ class MessageComposer:
                 role = match.group("role")
                 content = match.group("content")
                 self.messages.append({"role": role, "content": content.strip()})
-        elif self.model in ["gpt-neo-2.7B"]:
-            pair_pattern = r"GPT4 Correct User:(?P<inst>[\s\S]*?)<\|end_of_turn\|>\s*GPT4 Correct Assistant:(?P<answer>[\s\S]*?)<\|end_of_turn\|>"
-            pair_matches = re.finditer(
-                pair_pattern, self.merged_str, flags=re.MULTILINE | re.IGNORECASE
+        elif self.model in ["zephyr-7b-beta"]:
+            pair_pattern = (
+                r"<s>\s*\[INST\](?P<inst>[\s\S]*?)\[/INST\](?P<answer>[\s\S]*?)</s>"
             )
+            pair_matches = re.finditer(pair_pattern, self.merged_str, re.MULTILINE)
             pair_matches_list = list(pair_matches)
+
             self.messages = self.convert_pair_matches_to_messages(pair_matches_list)
-            inst_pattern = r"GPT4 Correct User:(?P<inst>[\s\S]*?)<\|end_of_turn\|>"
-            inst_matches = re.finditer(
-                inst_pattern, self.merged_str, flags=re.MULTILINE | re.IGNORECASE
-            )
+
+            inst_pattern = r"\[INST\](?P<inst>[\s\S]*?)\[/INST\]"
+            inst_matches = re.finditer(inst_pattern, self.merged_str, re.MULTILINE)
             inst_matches_list = list(inst_matches)
+
             self.append_last_instruction_to_messages(
                 inst_matches_list, pair_matches_list
             )

networks/message_streamer.py

@@ -12,8 +12,7 @@ class MessageStreamer:
         "mixtral-8x7b": "mistralai/Mixtral-8x7B-Instruct-v0.1",  # 72.62, fast [Recommended]
         "mistral-7b": "mistralai/Mistral-7B-Instruct-v0.2",  # 65.71, fast
         "nous-mixtral-8x7b": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO",
-        "gpt-neo-2.7B": "EleutherAI/gpt-neo-2.7B",  # 68.89, fast
-        # "zephyr-7b-beta": "HuggingFaceH4/zephyr-7b-beta",  # ❌ Too Slow
+        "zephyr-7b-beta": "HuggingFaceH4/zephyr-7b-beta",  # ❌ Too Slow
         # "llama-70b": "meta-llama/Llama-2-70b-chat-hf",  # ❌ Require Pro User
         # "codellama-34b": "codellama/CodeLlama-34b-Instruct-hf",  # ❌ Low Score
         # "falcon-180b": "tiiuae/falcon-180B-chat",  # ❌ Require Pro User
@@ -23,15 +22,14 @@ class MessageStreamer:
         "mixtral-8x7b": "</s>",
         "mistral-7b": "</s>",
         "nous-mixtral-8x7b": "<|im_end|>",
-        "gpt-neo-2.7B": "<|end_of_turn|>",
+        "zephyr-7b-beta": "<|end_of_turn|>",
 
     }
     TOKEN_LIMIT_MAP = {
         "mixtral-8x7b": 32768,
         "mistral-7b": 32768,
         "nous-mixtral-8x7b": 32768,
-        "gpt-neo-2.7B": 2048,
-
+        "zephyr-7b-beta": 4096,
     }
     TOKEN_RESERVED = 100