Skip to content

armgit / camel

Public Imported from GitHub
0

🐫 CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org

feat: New open source models (#649)

Zhiyu Wang8082e80a26de
diff --git a/camel/types/enums.py b/camel/types/enums.py
index 7d00e88c8..f20982553 100644
--- a/camel/types/enums.py
+++ b/camel/types/enums.py
@@ -30,15 +30,19 @@ class ModelType(Enum):
GPT_4_TURBO = "gpt-4-turbo"
GPT_4O = "gpt-4o"
GLM_4 = "glm-4"
+ GLM_4_OPEN_SOURCE = "glm-4-open-source"
GLM_4V = 'glm-4v'
GLM_3_TURBO = "glm-3-turbo"
STUB = "stub"
LLAMA_2 = "llama-2"
+ LLAMA_3 = "llama-3"
VICUNA = "vicuna"
VICUNA_16K = "vicuna-16k"
+ QWEN_2 = "qwen-2"
+
# Legacy anthropic models
# NOTE: anthropic lagecy models only Claude 2.1 has system prompt support
CLAUDE_2_1 = "claude-2.1"
@@ -87,6 +91,9 @@ class ModelType(Enum):
r"""Returns whether this type of models is open-source."""
return self in {
ModelType.LLAMA_2,
+ ModelType.LLAMA_3,
+ ModelType.QWEN_2,
+ ModelType.GLM_4_OPEN_SOURCE,
ModelType.VICUNA,
ModelType.VICUNA_16K,
}
@@ -135,7 +142,7 @@ class ModelType(Enum):
return 128000
elif self is ModelType.GPT_4O:
return 128000
- elif self == ModelType.GLM_4:
+ elif self == ModelType.GLM_4_OPEN_SOURCE:
return 8192
elif self == ModelType.GLM_3_TURBO:
return 8192
@@ -145,6 +152,12 @@ class ModelType(Enum):
return 4096
elif self is ModelType.LLAMA_2:
return 4096
+ elif self is ModelType.LLAMA_3:
+ return 8192
+ elif self is ModelType.QWEN_2:
+ return 128000
+ elif self is ModelType.GLM_4:
+ return 8192
elif self is ModelType.VICUNA:
# reference: https://lmsys.org/blog/2023-03-30-vicuna/
return 2048
@@ -184,6 +197,20 @@ class ModelType(Enum):
self.value in model_name.lower()
or "llama2" in model_name.lower()
)
+ elif self is ModelType.LLAMA_3:
+ return (
+ self.value in model_name.lower()
+ or "llama3" in model_name.lower()
+ )
+ elif self is ModelType.QWEN_2:
+ return (
+ self.value in model_name.lower()
+ or "qwen2" in model_name.lower()
+ )
+ elif self is ModelType.GLM_4_OPEN_SOURCE:
+ return (
+ 'glm-4' in model_name.lower() or "glm4" in model_name.lower()
+ )
else:
return self.value in model_name.lower()
diff --git a/camel/utils/token_counting.py b/camel/utils/token_counting.py
index ed7719418..49a4df0d8 100644
--- a/camel/utils/token_counting.py
+++ b/camel/utils/token_counting.py
@@ -51,7 +51,7 @@ def messages_to_prompt(messages: List[OpenAIMessage], model: ModelType) -> str:
system_message = messages[0]["content"]
ret: str
- if model == ModelType.LLAMA_2:
+ if model == ModelType.LLAMA_2 or model == ModelType.LLAMA_3:
# reference: https://github.com/facebookresearch/llama/blob/cfc3fc8c1968d390eb830e65c63865e980873a06/llama/generation.py#L212
seps = [" ", " </s><s>"]
role_map = {"user": "[INST]", "assistant": "[/INST]"}
@@ -93,6 +93,45 @@ def messages_to_prompt(messages: List[OpenAIMessage], model: ModelType) -> str:
else:
ret += role + ":"
return ret
+ elif model == ModelType.GLM_4_OPEN_SOURCE:
+ system_prompt = f"[gMASK]<sop><|system|>\n{system_message}"
+ ret = system_prompt
+ for msg in messages[1:]:
+ role = msg["role"]
+ content = msg["content"]
+ if not isinstance(content, str):
+ raise ValueError(
+ "Currently multimodal context is not "
+ "supported by the token counter."
+ )
+ if content:
+ ret += "<|" + role + "|>" + "\n" + content
+ else:
+ ret += "<|" + role + "|>" + "\n"
+ return ret
+ elif model == ModelType.QWEN_2:
+ system_prompt = f"<|im_start|>system\n{system_message}<|im_end|>"
+ ret = system_prompt + "\n"
+ for msg in messages[1:]:
+ role = msg["role"]
+ content = msg["content"]
+ if not isinstance(content, str):
+ raise ValueError(
+ "Currently multimodal context is not "
+ "supported by the token counter."
+ )
+ if content:
+ ret += (
+ '<|im_start|>'
+ + role
+ + '\n'
+ + content
+ + '<|im_end|>'
+ + '\n'
+ )
+ else:
+ ret += '<|im_start|>' + role + '\n'
+ return ret
else:
raise ValueError(f"Invalid model type: {model}")