Vihindi-K/ConvoDrift_human_eval_conversational_dataset

ConvoDrift Human Evaluation Data

This folder contains human evaluations of ConvoDrift conversations from three annotators. Each conversation has six prompt-response pairs, quality ratings for up to Q1-Q8, and optional corrections to drift and direction labels.

Files

Configuration	Records	Description
annotator_1	500	Complete evaluation file for Annotator 1
annotator_2	500	Complete evaluation file for Annotator 2
annotator_3	500	Complete evaluation file for Annotator 3
common_annotator_1	175	Shared agreement set evaluated by Annotator 1
common_annotator_2	175	Shared agreement set evaluated by Annotator 2
common_annotator_3	175	Shared agreement set evaluated by Annotator 3

The complete files contain the same 500 conversation IDs. The three common_annotator_* files contain the same 175-ID evaluation subset used for inter-annotator agreement analysis.

Main Fields

• data.conversation_id: conversation identifier.

• data.communication_genre or data.domain: genre or domain metadata,

  depending on the annotator file.

• data.scenario: scenario represented by the conversation.

• data.pairs: six prompts and responses with original drift/direction labels.

• data.ratings: human ratings for criteria Q1-Q8.

• data.refined_drift_label: optional human corrections to drift labels.

• data.refined_direction_label: optional human corrections to direction labels.

A null value in a refined-label array means that no correction was made. In that case, use the original label from the corresponding item in data.pairs.