From 3f79f1d7121c0bab3b05632058eaf51d75fb7f21 Mon Sep 17 00:00:00 2001
From: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
Date: Mon, 2 Oct 2023 01:58:42 -0400
Subject: [PATCH] fix: Update Llama2 Formatting Template
- Modify _system_template to better align system messages
- Remove [INST] tag from system message format
This change aims to fix the repetitive greeting issue in the conversational context by refining the message template used for Llama2.
---
llama_cpp/llama_chat_format.py | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py
index 9a09a28ee..8774d05f8 100644
--- a/llama_cpp/llama_chat_format.py
+++ b/llama_cpp/llama_chat_format.py
@@ -1,5 +1,6 @@
import dataclasses
-from typing import Any, Dict, List, Optional, Tuple, Union, Protocol
+from typing import Any, Dict, List, Optional, Protocol, Tuple, Union
+
from . import llama_types
@@ -144,7 +145,7 @@ def format_llama2(
messages: List[llama_types.ChatCompletionRequestMessage],
**kwargs: Any,
) -> ChatFormatterResponse:
- _system_template = "[INST] <>\n{system_message}\n<>\n\n"
+ _system_template = "<>\n{system_message}\n<>\n\n"
_roles = dict(user="[INST]", assistant="[/INST]")
_sep = "\n\n"
system_message = _get_system_message(messages)