Implementation:Deepseek ai Janus Conversation Template

Knowledge Sources	Janus
Domains	NLP, Prompt_Engineering
Last Updated	2026-02-10 09:30 GMT

Overview

Concrete tool for managing conversation templates and rendering multi-turn prompts provided by the Janus conversation module.

Description

The Conversation dataclass and get_conv_template function provide a template registry system for formatting multi-turn dialogues. The Conversation class stores template metadata (role names, separators, stop tokens) and conversation history, then renders the full prompt via get_prompt(). The "deepseek" template is registered by default with <|User|> / <|Assistant|> role markers.

Usage

Import get_conv_template when you need to manually construct prompts from conversation messages. In practice, the VLChatProcessor.apply_sft_template_for_multi_turn_prompts method wraps this functionality, so direct usage is less common.

Code Reference

Source Location

• Repository: Janus
• File: janus/utils/conversation.py
• Lines: L52-237 (Conversation class), L235-237 (get_conv_template)

Signature

@dataclasses.dataclass
class Conversation:
    name: str
    system_template: str = "{system_message}"
    system_message: str = ""
    roles: List[str] = (("USER", "ASSISTANT"),)
    messages: List[List[str]] = ()
    offset: int = 0
    sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE
    sep: str = "\n"
    sep2: str = None
    stop_str: str = None
    stop_token_ids: List[int] = None

    def get_prompt(self) -> str: ...
    def append_message(self, role: str, message: str): ...
    def copy(self) -> "Conversation": ...

def get_conv_template(name: str) -> Conversation:
    """Get a conversation template by name (returns a copy)."""

Import

from janus.utils.conversation import Conversation, get_conv_template

I/O Contract

Inputs

Name	Type	Required	Description
name	str	Yes	Template name (e.g., "deepseek", "llama-2", "plain")

Outputs

Name	Type	Description
conversation	Conversation	Conversation object with template settings and empty message history
get_prompt() returns	str	Fully formatted prompt string with role markers and separators

Usage Examples

Basic Prompt Construction

from janus.utils.conversation import get_conv_template

# Get the DeepSeek conversation template
conv = get_conv_template("deepseek")

# Add messages
conv.append_message(conv.roles[0], "What is in this image?")
conv.append_message(conv.roles[1], "")  # Empty = model should generate

# Render the formatted prompt
prompt = conv.get_prompt()
# Output: "<|User|>: What is in this image?\n\n<|Assistant|>:"

With System Prompt

conv = get_conv_template("deepseek")
conv.set_system_message("You are a helpful visual assistant.")
conv.append_message(conv.roles[0], "Describe the scene.")
conv.append_message(conv.roles[1], "")

prompt = conv.get_prompt()
# Output: "You are a helpful visual assistant.\n\n<|User|>: Describe the scene.\n\n<|Assistant|>:"

Related Pages

Implements Principle

• Principle:Deepseek_ai_Janus_Conversation_Formatting