Auditing (show_audit_settings = !show_audit_settings)} > help_outline

In this section, we'll be auditing the content moderation system. Here, you’ll be aided by a personalized model that will help direct your attention towards potential problem areas in the model’s performance. This model isn’t meant to be perfect, but is designed to help you better focus on areas that need human review.

{#if show_audit_settings}

Audit settings

Choose your audit settings here. These settings will affect all of the visualizations that follow, so you can return back here to make changes.

{#key personalized_model}

{/key}

Current model: {personalized_model}

{/if}

{#if audit_type == audit_types[0]}

Overview of all topics

First, browse the system performance by different auto-generated comment topic areas.

{#await promise}

{:then audit_results} {#if audit_results} {/if} {:catch error}

{error.message}

{/await}

{/if} {#if show_topic_training}

Topic model training

In what topic area would you like to tune your model?

Comments to label

Comments with scores 0 and 1 will be allowed to remain on the platform.
Comments with scores 2, 3, or 4 will be deleted from the platform.
Given that some comments may lack context, if you're not sure, feel free to mark the unsure option to skip a comment.

{#key topic} {/key}

{/if}

Topic exploration

What topic would you like to explore further?

{#if use_model}

What kind of system errors do you want to focus on?

{#each error_type_options as e}

{e.opt} {e.descr} help_outline

{/each}

{/if}

{#await promise_cluster}

{:then cluster_results} {#if cluster_results} {#if topic} {/if} {/if} {:catch error}

{error.message}

{/await}

Gather additional evidence

Next, you can optionally search for more comments to serve as evidence through manual keyword search (for individual words or phrases).

{#key error_type} {/key}

Finalize your current report

Finally, review the report you've generated on the side panel and provide a brief summary of the problem you see. You may also list suggestions or insights into addressing this problem if you have ideas. This report will be directly used by the model developers to address the issue you've raised