integrations.swanlab.completion_logger

integrations.swanlab.completion_logger

SwanLab completion logger for RLHF/DPO/KTO/ORPO/GRPO training.

This module provides utilities for logging model completions during preference training to SwanLab for qualitative analysis.

Classes

Name	Description
CompletionLogger	Memory-bounded logger for RLHF completions.

CompletionLogger

integrations.swanlab.completion_logger.CompletionLogger(maxlen=128)

Memory-bounded logger for RLHF completions.

Stores prompts, completions, and rewards in fixed-size deques to prevent memory leaks during long training runs. Logs completion tables to SwanLab for qualitative analysis of model outputs.

Example usage

logger = CompletionLogger(maxlen=128) logger.add_dpo_completion( … step=0, … prompt=“What is AI?”, … chosen=“Artificial Intelligence is…”, … rejected=“AI means…”, … reward_diff=0.5 … ) logger.log_to_swanlab()

Attributes

Name	Type	Description
maxlen		Maximum number of completions to store (older ones are dropped)
data	deque[Mapping[str, Any]]	Deque storing completion dictionaries

Methods

Name	Description
add_dpo_completion	Add a DPO completion to the buffer.
add_grpo_completion	Add a GRPO completion to the buffer.
add_kto_completion	Add a KTO completion to the buffer.
add_orpo_completion	Add an ORPO completion to the buffer.
clear	Clear all buffered completions.
log_to_swanlab	Log buffered completions to SwanLab as a table.

add_dpo_completion

integrations.swanlab.completion_logger.CompletionLogger.add_dpo_completion(
    step,
    prompt,
    chosen,
    rejected,
    reward_diff=None,
)

Add a DPO completion to the buffer.

Parameters

Name	Type	Description	Default
step	int	Training step number	required
prompt	str	Input prompt	required
chosen	str	Chosen (preferred) completion	required
rejected	str	Rejected (non-preferred) completion	required
reward_diff	float | None	Reward difference (chosen - rejected), if available	None

add_grpo_completion

integrations.swanlab.completion_logger.CompletionLogger.add_grpo_completion(
    step,
    prompt,
    completion,
    reward=None,
    advantage=None,
)

Add a GRPO completion to the buffer.

Parameters

Name	Type	Description	Default
step	int	Training step number	required
prompt	str	Input prompt	required
completion	str	Model-generated completion	required
reward	float | None	Reward score from reward model	None
advantage	float | None	Advantage estimate (reward - baseline)	None

add_kto_completion

integrations.swanlab.completion_logger.CompletionLogger.add_kto_completion(
    step,
    prompt,
    completion,
    label,
    reward=None,
)

Add a KTO completion to the buffer.

Parameters

Name	Type	Description	Default
step	int	Training step number	required
prompt	str	Input prompt	required
completion	str	Model-generated completion	required
label	bool	True if desirable, False if undesirable	required
reward	float | None	Reward score, if available	None

add_orpo_completion

integrations.swanlab.completion_logger.CompletionLogger.add_orpo_completion(
    step,
    prompt,
    chosen,
    rejected,
    log_odds_ratio=None,
)

Add an ORPO completion to the buffer.

Parameters

Name	Type	Description	Default
step	int	Training step number	required
prompt	str	Input prompt	required
chosen	str	Chosen (preferred) completion	required
rejected	str	Rejected (non-preferred) completion	required
log_odds_ratio	float | None	Log odds ratio between chosen and rejected	None

clear

integrations.swanlab.completion_logger.CompletionLogger.clear()

Clear all buffered completions.

log_to_swanlab

integrations.swanlab.completion_logger.CompletionLogger.log_to_swanlab(
    table_name='completions',
)

Log buffered completions to SwanLab as a table.

Creates a SwanLab echarts Table with all buffered completions. Only logs if SwanLab is initialized and data is available.

Parameters

Name	Type	Description	Default
table_name	str	Name of the table in SwanLab dashboard. Default: “completions”	'completions'

Returns

Name	Type	Description
	bool	True if logging succeeded, False otherwise