nemo_rl.data.deepseek_v4_encoding

DeepSeek-V4 Encoding

A self-contained implementation for encoding/decoding DeepSeek-V4 chat messages with tool calling, thinking mode, and quick instruction task support.

Module Contents

Functions

to_json	Serialize a value to JSON string.
tools_from_openai_format	Extract function definitions from OpenAI-format tool list.
tool_calls_from_openai_format	Convert OpenAI-format tool calls to internal format.
tool_calls_to_openai_format	Convert internal tool calls to OpenAI format.
encode_arguments_to_dsml	Encode tool call arguments into DSML parameter format.
decode_dsml_to_arguments	Decode DSML parameters back to a tool call dict.
render_tools	Render tool schemas into the system prompt format.
find_last_user_index	Find the index of the last user/developer message.
render_message	Render a single message at the given index into its encoded string form.
merge_tool_messages	Merge tool messages into the preceding user message using content_blocks format.
sort_tool_results_by_call_order	Sort tool_result blocks within user messages by the order of tool_calls in the preceding assistant message.
encode_messages	Encode a list of messages into the DeepSeek-V4 prompt format.
_drop_thinking_messages	Drop reasoning and non-essential messages before the last user message.
_read_until_stop	Read text from index until one of the stop strings is found.
parse_tool_calls	Parse DSML tool calls from text starting at the given index.
parse_message_from_completion_text	Parse a model completion text into a structured assistant message.

Data

bos_token
eos_token
thinking_start_token
thinking_end_token
dsml_token
USER_SP_TOKEN
ASSISTANT_SP_TOKEN
LATEST_REMINDER_SP_TOKEN
DS_TASK_SP_TOKENS
VALID_TASKS
system_msg_template
user_msg_template
latest_reminder_msg_template
assistant_msg_template
assistant_msg_wo_eos_template
thinking_template
response_format_template
tool_call_template
tool_calls_template
tool_calls_block_name
tool_output_template
REASONING_EFFORT_MAX
TOOLS_TEMPLATE

API

nemo_rl.data.deepseek_v4_encoding.bos_token: str

‘<｜begin▁of▁sentence｜>’

nemo_rl.data.deepseek_v4_encoding.eos_token: str

‘<｜end▁of▁sentence｜>’

nemo_rl.data.deepseek_v4_encoding.thinking_start_token: str

‘’

nemo_rl.data.deepseek_v4_encoding.thinking_end_token: str

‘’

nemo_rl.data.deepseek_v4_encoding.dsml_token: str

‘｜DSML｜’

nemo_rl.data.deepseek_v4_encoding.USER_SP_TOKEN

‘<｜User｜>’

nemo_rl.data.deepseek_v4_encoding.ASSISTANT_SP_TOKEN

‘<｜Assistant｜>’

nemo_rl.data.deepseek_v4_encoding.LATEST_REMINDER_SP_TOKEN

‘<｜latest_reminder｜>’

nemo_rl.data.deepseek_v4_encoding.DS_TASK_SP_TOKENS

None

nemo_rl.data.deepseek_v4_encoding.VALID_TASKS

‘set(…)’

nemo_rl.data.deepseek_v4_encoding.system_msg_template: str

‘{content}’

nemo_rl.data.deepseek_v4_encoding.user_msg_template: str

‘{content}’

nemo_rl.data.deepseek_v4_encoding.latest_reminder_msg_template: str

‘{content}’

nemo_rl.data.deepseek_v4_encoding.assistant_msg_template: str

None

nemo_rl.data.deepseek_v4_encoding.assistant_msg_wo_eos_template: str

‘{reasoning}{content}{tool_calls}’

nemo_rl.data.deepseek_v4_encoding.thinking_template: str

‘{reasoning}’

nemo_rl.data.deepseek_v4_encoding.response_format_template: str = <Multiline-String>

nemo_rl.data.deepseek_v4_encoding.tool_call_template: str = <Multiline-String>

nemo_rl.data.deepseek_v4_encoding.tool_calls_template = <Multiline-String>

nemo_rl.data.deepseek_v4_encoding.tool_calls_block_name: str

‘tool_calls’

nemo_rl.data.deepseek_v4_encoding.tool_output_template: str

‘<tool_result>{content}</tool_result>’

nemo_rl.data.deepseek_v4_encoding.REASONING_EFFORT_MAX = <Multiline-String>

nemo_rl.data.deepseek_v4_encoding.TOOLS_TEMPLATE = <Multiline-String>

nemo_rl.data.deepseek_v4_encoding.to_json(value: Any) → str

Serialize a value to JSON string.

nemo_rl.data.deepseek_v4_encoding.tools_from_openai_format(tools)

Extract function definitions from OpenAI-format tool list.

nemo_rl.data.deepseek_v4_encoding.tool_calls_from_openai_format(tool_calls)

Convert OpenAI-format tool calls to internal format.

nemo_rl.data.deepseek_v4_encoding.tool_calls_to_openai_format(tool_calls)

Convert internal tool calls to OpenAI format.

nemo_rl.data.deepseek_v4_encoding.encode_arguments_to_dsml(

tool_call: Dict[str, Any],

) → str

Encode tool call arguments into DSML parameter format.

Parameters:

tool_call – Dict with “name” and “arguments” keys.

Returns:

DSML-formatted parameter string.

nemo_rl.data.deepseek_v4_encoding.decode_dsml_to_arguments(

tool_name: str,

tool_args: Dict[str, Tuple[str, str]],

) → Dict[str, str]

Decode DSML parameters back to a tool call dict.

Parameters:

• tool_name – Name of the tool.

• tool_args – Dict mapping param_name -> (value, is_string_flag).

Returns:

Dict with “name” and “arguments” (JSON string) keys.

nemo_rl.data.deepseek_v4_encoding.render_tools(

tools: List[Dict[str, Union[str, Dict[str, Any]]]],

) → str

Render tool schemas into the system prompt format.

Parameters:

tools – List of tool schema dicts (each with name, description, parameters).

Returns:

Formatted tools section string.

nemo_rl.data.deepseek_v4_encoding.find_last_user_index(

messages: List[Dict[str, Any]],

) → int

Find the index of the last user/developer message.

nemo_rl.data.deepseek_v4_encoding.render_message(

index: int,

messages: List[Dict[str, Any]],

thinking_mode: str,

drop_thinking: bool = True,

reasoning_effort: Optional[str] = None,

) → str

Render a single message at the given index into its encoded string form.

This is the core function that converts each message in the conversation into the DeepSeek-V4 format.

Parameters:

• index – Index of the message to render.

• messages – Full list of messages in the conversation.

• thinking_mode – Either “chat” or “thinking”.

• drop_thinking – Whether to drop reasoning content from earlier turns.

• reasoning_effort – Optional reasoning effort level (“max”, “high”, or None).

Returns:

Encoded string for this message.

nemo_rl.data.deepseek_v4_encoding.merge_tool_messages(

messages: List[Dict[str, Any]],

) → List[Dict[str, Any]]

Merge tool messages into the preceding user message using content_blocks format.

DeepSeek-V4 does not have a standalone “tool” role; instead, tool results are encoded as <tool_result> blocks within user messages.

This function converts a standard OpenAI-format conversation (with separate “tool” role messages) into V4 format where tool results are merged into user messages.

Parameters:

messages – List of message dicts in OpenAI format.

Returns:

Processed message list with tool messages merged into user messages.

nemo_rl.data.deepseek_v4_encoding.sort_tool_results_by_call_order(

messages: List[Dict[str, Any]],

) → List[Dict[str, Any]]

Sort tool_result blocks within user messages by the order of tool_calls in the preceding assistant message.

Parameters:

messages – Preprocessed message list (after merge_tool_messages).

Returns:

Message list with sorted tool result blocks.

nemo_rl.data.deepseek_v4_encoding.encode_messages(

messages: List[Dict[str, Any]],

thinking_mode: str,

context: Optional[List[Dict[str, Any]]] = None,

drop_thinking: bool = True,

add_default_bos_token: bool = True,

reasoning_effort: Optional[str] = None,

) → str

Encode a list of messages into the DeepSeek-V4 prompt format.

This is the main entry point for encoding conversations. It handles:

• BOS token insertion

• Thinking mode with optional reasoning content dropping

• Tool message merging into user messages

• Multi-turn conversation context

Parameters:

• messages – List of message dicts to encode.

• thinking_mode – Either “chat” or “thinking”.

• context – Optional preceding context messages (already encoded prefix).

• drop_thinking – If True, drop reasoning from earlier assistant turns (only keep reasoning for messages after the last user message).

• add_default_bos_token – Whether to prepend BOS token at conversation start.

• reasoning_effort – Optional reasoning effort level (“max”, “high”, or None).

Returns:

The encoded prompt string.

nemo_rl.data.deepseek_v4_encoding._drop_thinking_messages(

messages: List[Dict[str, Any]],

) → List[Dict[str, Any]]

Drop reasoning and non-essential messages before the last user message.

Behavior:

• Messages with role in [“user”, “system”, “tool”, “latest_reminder”] are always kept.

• Messages at or after the last user index are always kept.

• Assistant messages before the last user get reasoning removed.

• Developer messages before the last user are dropped entirely.

nemo_rl.data.deepseek_v4_encoding._read_until_stop(

index: int,

text: str,

stop: List[str],

) → Tuple[int, str, Optional[str]]

Read text from index until one of the stop strings is found.

Returns:

Tuple of (new_index, content_before_stop, matched_stop_string_or_None).

nemo_rl.data.deepseek_v4_encoding.parse_tool_calls(

index: int,

text: str,

) → Tuple[int, Optional[str], List[Dict[str, str]]]

Parse DSML tool calls from text starting at the given index.

Parameters:

• index – Starting position in text.

• text – The full text to parse.

Returns:

Tuple of (new_index, last_stop_token, list_of_tool_call_dicts). Each tool call dict has “name” and “arguments” keys.

nemo_rl.data.deepseek_v4_encoding.parse_message_from_completion_text(

text: str,

thinking_mode: str,

) → Dict[str, Any]

Parse a model completion text into a structured assistant message.

This function takes the raw text output from the model (a single assistant turn) and extracts:

• reasoning (thinking block)

• content (summary/response)

• tool_calls (if any)

NOTE: This function is designed to parse only correctly formatted strings and will raise ValueError for malformed output.

Parameters:

• text – The raw completion text (including EOS token).

• thinking_mode – Either “chat” or “thinking”.

Returns:

“role”, “content”, “reasoning”, “tool_calls”. tool_calls are in OpenAI format.

Return type:

Dict with keys