pavidu/PromptAnonBench

Dataset used by our paper Can Prompt Anonymity Protect Your Identity From LLM Providers?.

Contains user prompts only of WildChat, SWE-Chat, and ShareChat, and various transformed variations, along with Gemini Embedding 2 for each conversation (user-prompt only, not assistant prompt).

Each file has the following name format: <wildchat|swe_chat|sharechat>[_defense][_feature].parquet

For example: wildchat_styleremix_gemini_embedding_2.parquet contains the Gemini Embeddding 2 of the StyleRemixed version of WildChat.

Explanation of column names (for wildchat.parquet, swe_chat.parquet, and sharechat.parquet):

• doc_id: Unique ID of a conversation. Has this format: <wc|sc|sh>-<original_id>. The first two characters correspond to WildChat, ShareChat, and ShareChat, respectively. <original_id> is copied from the original datasets, so you can use this part to find the original full conversations.
• source: Name of source dataset.
• author_id: Assigned unique ID of an author. (Not defined for ShareChat.) Has this format: <wildchat|swe-chat>-<hash>.
• turns: List of user prompts in the conversation, ordered by turn. Note that these are slightly different from the originals, since we further preprocessed the user prompts to get rid of more URL, file paths, etc.
• num_turns: Number of user prompts.
• language_primary: Primary language detected by lingua.
• language_secondary: Secondary language detected by lingua.
• model: The LLM used in the conversation.
• model_owner: The name of the company that provides the LLM (e.g., OpenAI, Anthropic).
• agent: (Defined only for SWE-Chat) The name of the agent used the conversation.
• started_at: UTC time of the first turn.
• ended_at: UTC time of the last turn.

If you find our dataset useful, please cite us:

@misc{pham2026promptanonymity,
      title={Can Prompt Anonymity Protect Your Identity From LLM Providers?}, 
      author={Dzung Pham and Dillon Sheils and Naina Singh and Amir Houmansadr},
      year={2026},
      eprint={2609.33903},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2609.33903}, 
}