Principle:Deepseek ai Janus Conversation Formatting
| Knowledge Sources | |
|---|---|
| Domains | NLP, Prompt_Engineering |
| Last Updated | 2026-02-10 09:30 GMT |
Overview
A template-based system for formatting multi-turn conversations into structured prompts that language models can process.
Description
Conversation formatting converts a list of message dictionaries (with roles like "User" and "Assistant") into a single string prompt that follows a specific template format. This is critical for instruction-tuned models that expect inputs in a particular supervised fine-tuning (SFT) format.
The Janus repository uses a conversation template system adapted from FastChat. Each template defines role names, separator tokens, and formatting rules. The primary template used is "deepseek", which uses <|User|> and <|Assistant|> role markers with <｜end▁of▁sentence｜> as the end-of-turn separator.
Usage
Use this principle whenever preparing text input for Janus models, whether for multimodal understanding or image generation. The conversation must be formatted into the correct SFT template before tokenization.
Theoretical Basis
The conversation template follows a structured format:
# Abstract template structure (DeepSeek style)
# {system_message}\n\n
# <|User|>: {user_message}\n\n
# <|Assistant|>: {assistant_message}<｜end▁of▁sentence｜>
# <|User|>: {next_message}\n\n
# <|Assistant|>:
Key design decisions:
- Role markers (<|User|>, <|Assistant|>) delimit turn boundaries
- Separator tokens (\n\n between roles, <｜end▁of▁sentence｜> after assistant) signal turn completion
- Empty assistant content signals the model should generate a response
- The template is registered globally and retrieved by name via get_conv_template