Files
exllamav3/examples/chat_templates.py
2026-04-05 19:44:46 +02:00

824 lines
26 KiB
Python

class PromptFormat:
def __init__(self, user_name, bot_name):
self.user_name = user_name
self.bot_name = bot_name
def set_special(self, spc: dict):
pass
def default_system_prompt(self, think):
raise NotImplementedError()
def format(self, system_prompt, messages, think):
raise NotImplementedError()
def add_bos(self):
raise NotImplementedError()
def thinktag(self):
return "<think>", "</think>"
class PromptFormat_raw(PromptFormat):
description = "Model-agnostic mode simulating a raw chatlog"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
f"This is a conversation between a helpful AI assistant " +
(f"named {self.bot_name} " if self.bot_name != "Assistant" else "") +
(f"and a user named {self.user_name}." if self.user_name != "User" else """and a user.""")
)
def format(self, system_prompt, messages, think):
context = system_prompt + "\n"
for (u, a) in messages:
context += f"{self.user_name}: {u}\n"
context += f"{self.bot_name}:"
if a is not None:
context += f"{a}\n"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
self.user_name + ":",
self.user_name[0:1] + ":",
self.user_name.upper() + ":",
self.user_name.lower() + ":",
tokenizer.eos_token_id
]
class PromptFormat_llama3(PromptFormat):
description = "Llama3-instruct models"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
"""Assist users with tasks and answer questions to the best of your knowledge. Provide helpful and informative """
"""responses. Be conversational and engaging. If you are unsure or lack knowledge on a topic, admit it and try """
"""to find the answer or suggest where to find it. Keep responses concise and relevant. Follow ethical """
"""guidelines and promote a safe and respectful interaction."""
)
def format(self, system_prompt, messages, think):
context = f"<|start_header_id|>system<|end_header_id|>\n\n{system_prompt}<|eot_id|>"
for (u, a) in messages:
context += f"<|start_header_id|>user<|end_header_id|>\n\n{u}<|eot_id|>"
context += f"<|start_header_id|>assistant<|end_header_id|>\n\n"
if a is not None: context += f"{a}<|eot_id|>"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|eot_id|>"),
tokenizer.single_id("<|start_header_id|>")
]
class PromptFormat_chatml(PromptFormat):
description = "ChatML format, as used by e.g. Qwen"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
f"You are a helpful AI assistant."
)
def format(self, system_prompt, messages, think):
context = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
for (u, a) in messages:
context += f"<|im_start|>user\n{u}<|im_end|>\n"
context += f"<|im_start|>assistant\n"
if a is not None: context += f"{a}<|im_end|>\n"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|im_end|>"),
"""<|im_end|>"""
]
class PromptFormat_qwen35(PromptFormat):
description = "Qwen3.5 format, reasoning-aware ChatML"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
f"You are a helpful AI assistant."
)
def format(self, system_prompt, messages, think):
context = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
for (u, a) in messages:
context += f"<|im_start|>user\n{u}<|im_end|>\n"
context += f"<|im_start|>assistant\n"
if a is not None:
context += f"{a}<|im_end|>\n"
elif not think:
context += f"<think>\n\n</think>\n\n"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|im_end|>"),
"""<|im_end|>"""
]
def thinktag(self):
return "<think>\n", "</think>"
class PromptFormat_phi(PromptFormat):
description = "Phi3/Phi4 instruct models (note, some Phi4 models use ChatML)"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
f"You are a helpful AI assistant."
)
def format(self, system_prompt, messages, think):
context = f"<|system|>\n{system_prompt}<|end|>\n"
for (u, a) in messages:
context += f"<|user|>\n{u}<|end|>\n"
context += f"<|assistant|>\n"
if a is not None: context += f"{a}<|end|>\n"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|end|>"),
]
class PromptFormat_glm(PromptFormat):
description = "ChatGLM(4) models"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
f"You are a helpful AI assistant."
)
def format(self, system_prompt, messages, think):
context = f"[gMASK]<sop><|system|>\n{system_prompt}"
for (u, a) in messages:
context += f"<|user|>\n{u}"
context += f"<|assistant|>"
if a is not None: context += f"{a}"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|user|>"),
tokenizer.single_id("</answer>"),
tokenizer.single_id("<|observation|>"),
]
class PromptFormat_mistral(PromptFormat):
description = "Mistral-instruct models (v3)"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
"""You are a helpful AI assistant."""
)
def format(self, system_prompt, messages, think):
context = ""
first = True
for (u, a) in messages:
if first:
context += f"[INST] {system_prompt}\n\n{u}[/INST]"
first = False
else:
context += f"[INST] {u}[/INST]"
if a is not None: context += f" {a}</s>"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id
]
class PromptFormat_mistral3(PromptFormat):
description = "Mistral3/Ministral-instruct models"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
"""You are a helpful AI assistant."""
)
def format(self, system_prompt, messages, think):
context = f"[SYSTEM_PROMPT]{system_prompt}[/SYSTEM_PROMPT]"
for (u, a) in messages:
context += f"[INST]{u}[/INST]"
if a is not None: context += f"{a}"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id
]
class PromptFormat_gemma(PromptFormat):
description = "Gemma"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return ""
def format(self, system_prompt, messages, think):
context = ""
for (u, a) in messages:
context += f"<start_of_turn>user\n"
context += f"{u}<end_of_turn>\n"
context += f"<start_of_turn>model\n"
if a is not None: context += f"{a}<end_of_turn>\n"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<end_of_turn>"),
tokenizer.single_id("<start_of_turn>"),
]
class PromptFormat_gemma4(PromptFormat):
description = "Gemma4"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
"""You are a helpful AI assistant."""
)
def format(self, system_prompt, messages, think):
context = ""
if system_prompt:
context = "<|turn>system\n"
if think:
context += "<|think|>"
context += system_prompt + "<turn|>\n"
for (u, a) in messages:
context += f"<|turn>user\n"
context += f"{u}<turn|>\n"
context += f"<|turn>model\n"
if a is not None: context += f"{a}<turn|>\n"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<eos>"),
tokenizer.single_id("<turn|>"),
]
def thinktag(self):
return "<|channel>thought", "<channel|>"
class PromptFormat_reka(PromptFormat):
description = "Reka Flash 3"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return ""
def format(self, system_prompt, messages, think):
context = ""
first = True
for (u, a) in messages:
if first and system_prompt:
context += f"human: {system_prompt} {u} <sep> "
first = False
else:
context += f"human: {u} <sep> "
context += f"assistant:"
if a is not None: context += f" {a} <sep> "
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
"<sep>",
]
def thinktag(self):
return " <reasoning>\n", "</reasoning>"
class PromptFormat_cohere(PromptFormat):
description = "Cohere"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
"## Task and Context\n"
"You help people answer their questions and other requests interactively. You will be asked a very "
"wide array of requests on all kinds of topics. You should focus on serving the user's needs as "
"best you can, which will be wide-ranging.\n\n"
"## Style Guide\n"
"Unless the user asks for a different style of answer, you should answer in full sentences, using "
"proper grammar and spelling."
)
def format(self, system_prompt, messages, think):
context = ""
if system_prompt:
context += "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>"
context += system_prompt
context += "<|END_OF_TURN_TOKEN|>"
for (u, a) in messages:
context += "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>"
context += u
context += "<|END_OF_TURN_TOKEN|>"
context += "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
if a is not None:
context += a
context += "<|END_OF_TURN_TOKEN|>"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
"<|END_OF_TURN_TOKEN|>",
]
class PromptFormat_dots(PromptFormat):
description = "Dots"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return "You are a helpful assistant."
def format(self, system_prompt, messages, think):
context = ""
if system_prompt:
context += "<|system|>"
context += system_prompt
context += "<|endofsystem|>"
for (u, a) in messages:
context += "<|userprompt|>"
context += u
context += "<|endofuserprompt|>"
context += "<|response|>"
if a is not None:
context += a
context += "<|endofresponse|>"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|endofresponse|>"),
"<|endofresponse|>",
]
def thinktag(self):
return "<|sec-cot|>", "<|sec-end-cot|>"
class PromptFormat_ernie(PromptFormat):
description = "Ernie"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return "You are a helpful assistant."
def format(self, system_prompt, messages, think):
context = "<|begin_of_sentence|>"
if system_prompt:
context += system_prompt
context += "\n"
for (u, a) in messages:
context += "User: "
context += u
context += "\n"
context += "Assistant: "
if a is not None:
context += a
context += "<|end_of_sentence|>"
return context
def add_bos(self):
return True
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|end_of_sentence|>")
]
class PromptFormat_smollm3(PromptFormat):
description = "SmolLM3 format (ChatML)"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
from datetime import datetime
today_str = datetime.today().strftime("%d %B %Y")
sp = (
"## Metadata\n"
"\n"
"Knowledge Cutoff Date: June 2025\n"
f"Today Date: {today_str}\n"
f"Reasoning Mode: {'/think' if think else '/no_think'}\n"
"\n"
"## Custom Instructions\n"
"\n"
)
if think:
sp += (
"You are a helpful AI assistant named SmolLM, trained by Hugging Face. Your role as an assistant involves "
"thoroughly exploring questions through a systematic thinking process before providing the final precise "
"and accurate solutions. This requires engaging in a comprehensive cycle of analysis, summarizing, "
"exploration, reassessment, reflection, backtracking, and iteration to develop well-considered thinking "
"process. Please structure your response into two main sections: Thought and Solution using the specified "
"format: <think> Thought section </think> Solution section. In the Thought section, detail your reasoning "
"process in steps. Each step should include detailed considerations such as analysing questions, "
"summarizing relevant findings, brainstorming new ideas, verifying the accuracy of the current steps, "
"refining any errors, and revisiting previous steps. In the Solution section, based on various attempts, "
"explorations, and reflections from the Thought section, systematically present the final solution that "
"you deem correct. The Solution section should be logical, accurate, and concise and detail necessary "
"steps needed to reach the conclusion.\n\n"
)
else:
sp += "You are a helpful AI assistant named SmolLM, trained by Hugging Face.\n\n"
return sp
def format(self, system_prompt, messages, think):
context = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
for (u, a) in messages:
context += f"<|im_start|>user\n{u}<|im_end|>\n"
context += f"<|im_start|>assistant\n"
if a is not None: context += f"{a}<|im_end|>\n"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|im_end|>"),
"""<|im_end|>"""
]
class PromptFormat_commanda(PromptFormat):
description = "Command-A (Cohere variant)"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
# Abridged version
return (
"You are a helpful AI assistant.\n"
"\n"
"Your information cutoff date is June 2024.\n"
"\n"
"You help people answer their questions and other requests interactively. You will be asked a very "
"wide array of requests on all kinds of topics. Focus on serving the user's needs as best you can, "
"which will be wide-ranging.\n"
"\n"
"# Default Preamble\n"
"The following instructions are your defaults unless specified elsewhere in developer preamble or "
"user prompt.\n"
"- Your name is Command.\n"
"- You are a large language model built by Cohere.\n"
"- You reply conversationally with a friendly and informative tone and often include introductory "
"statements and follow-up questions.\n"
"- If the input is ambiguous, ask clarifying follow-up questions.\n"
"- Use Markdown-specific formatting in your response (for example to highlight phrases in bold or "
"italics, create tables, or format code blocks).\n"
"- Avoid using LaTeX to generate mathematical notation for complex equations.\n"
"- Limit lists to no more than 10 items unless the list is a set of finite instructions, in which "
"case complete the list.\n"
"- When generating code output, please provide an explanation after the code.\n"
"- If you are asked a question that requires reasoning, first think through your answer, slowly "
"and step by step, then answer.\n"
)
def format(self, system_prompt, messages, think):
context = ""
if system_prompt:
context += "<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>"
context += system_prompt
context += "<|END_OF_TURN_TOKEN|>"
for (u, a) in messages:
context += "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>"
context += u
context += "<|END_OF_TURN_TOKEN|>"
context += "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><|START_RESPONSE|>"
if a is not None:
context += a
context += "<|END_RESPONSE|><|END_OF_TURN_TOKEN|>"
return context
def add_bos(self):
return True # HF template uses double BOS token
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|END_RESPONSE|>"),
tokenizer.single_id("<|END_OF_TURN_TOKEN|>"),
"<|END_OF_TURN_TOKEN|>",
]
class PromptFormat_exaone(PromptFormat):
description = "EXAONE (EXAONE4)"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
# Abridged version
return (
"You are a helpful AI assistant."
)
def format(self, system_prompt, messages, think):
context = ""
if system_prompt:
context += "[|system|]\n"
context += system_prompt
context += "[|endofturn|]\n"
for (u, a) in messages:
context += "[|user|]\n"
context += u
context += "[|endofturn|]\n"
context += "[|assistant|]\n"
if a is not None:
context += a
context += "[|endofturn|]\n"
return context
def add_bos(self):
return False # HF template uses double BOS token
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("[|endofturn|]"),
"[|endofturn|]",
]
class PromptFormat_seed(PromptFormat):
description = "Seed-OSS"
def __init__(self, *args):
super().__init__(*args)
self.thinking_budget = 1024
def set_special(self, spc: dict):
self.thinking_budget = spc.get("thinking_budget", self.thinking_budget)
def default_system_prompt(self, think):
if think:
budget_reflections_v05 = {
0: 0,
512: 128,
1024: 256,
2048: 512,
4096: 512,
8192: 1024,
1e20: 1024
}
ns_interval = budget_reflections_v05[min([k for k in budget_reflections_v05 if k >= self.thinking_budget])]
return (
f"You are an intelligent assistant with reflective ability. In the process of thinking and reasoning, "
f"you need to strictly follow the thinking budget, which is {self.thinking_budget}. That is, you need to "
f"complete your thinking within {self.thinking_budget} tokens and start answering the user's questions. "
f"You will reflect on your thinking process every {ns_interval} tokens, stating how many tokens have "
f"been used and how many are left."
)
else:
return (
"You are an intelligent assistant that can answer questions in one step without the need for reasoning "
"and thinking, that is, your thinking budget is 0. Next, please skip the thinking process and directly "
"start answering the user's questions."
)
def format(self, system_prompt, messages, think):
context = ""
if system_prompt:
context += "<seed:bos>system\n"
context += system_prompt
context += "<seed:eos>"
for (u, a) in messages:
context += "<seed:bos>user\n"
context += u
context += "<seed:eos>"
context += "<seed:bos>assistant\n"
if a is not None:
context += a
context += "<seed:eos>"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<seed:eos>"),
"<seed:eos>",
]
def thinktag(self):
return "<seed:think>", "</seed:think>"
class PromptFormat_apertus(PromptFormat):
description = "Apertus"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
from datetime import datetime
today_str = datetime.today().strftime("%Y-%m-%d")
return (
f"You are Apertus, a helpful assistant created by the SwissAI initiative.\n"
f"Knowledge cutoff: 2024-04\n"
f"Current date: {today_str}"
)
def format(self, system_prompt, messages, think):
context = "<s>"
if system_prompt:
context += "<|system_start|>"
context += system_prompt
context += "<|system_end|>"
context += "<|developer_start|>Deliberation: "
context += "enabled" if think else "disabled"
context += "\nTool Capabilities: disabled<|developer_end|>"
for (u, a) in messages:
context += "<|user_start|>"
context += u
context += "<|user_end|>"
context += "<|assistant_start|>"
if a is not None:
context += a
context += "<|assistant_end|>"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("<|assistant_end|>"),
"<|assistant_end|>",
]
def thinktag(self):
return None, None
class PromptFormat_minimax(PromptFormat):
description = "MiniMax"
def __init__(self, *args):
super().__init__(*args)
def default_system_prompt(self, think):
return (
f"You are a helpful assistant."
)
def format(self, system_prompt, messages, think):
context = "]~!b["
if system_prompt:
context += "]~b]system\n"
context += system_prompt
context += "[e~[\n"
for (u, a) in messages:
context += "]~b]user\n"
context += u
context += "[e~[\n"
context += "]~b]ai\n"
if a is not None:
context += a
context += "[e~[\n"
return context
def add_bos(self):
return False
def stop_conditions(self, tokenizer):
return [
tokenizer.eos_token_id,
tokenizer.single_id("[e~["),
"[e~[",
]
def thinktag(self):
return "<think>", "</think>"
prompt_formats = {
"raw": PromptFormat_raw,
"llama3": PromptFormat_llama3,
"chatml": PromptFormat_chatml,
"qwen35": PromptFormat_qwen35,
"phi": PromptFormat_phi,
"mistral": PromptFormat_mistral,
"mistral3": PromptFormat_mistral3,
"gemma": PromptFormat_gemma,
"gemma4": PromptFormat_gemma4,
"glm": PromptFormat_glm,
"reka": PromptFormat_reka,
"cohere": PromptFormat_cohere,
"dots": PromptFormat_dots,
"ernie": PromptFormat_ernie,
"smollm3": PromptFormat_smollm3,
"commanda": PromptFormat_commanda,
"exaone": PromptFormat_exaone,
"seed": PromptFormat_seed,
"apertus": PromptFormat_apertus,
"minimax": PromptFormat_minimax,
}