| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: Molt-Intent |
| size_categories: |
| - 1K<n<10K |
| task_categories: |
| - text-classification |
| tags: |
| - content-moderation |
| - llm-agents |
| - agent-social-networks |
| - moltbook |
| - intent-detection |
| - synthetic |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train.parquet |
| - split: id |
| path: data/id.parquet |
| - split: id_spam |
| path: data/id_spam.parquet |
| - split: ood |
| path: data/ood.parquet |
| - split: ood_spam |
| path: data/ood_spam.parquet |
| --- |
| |
| # Molt-Intent |
|
|
| Molt-Intent is a dataset for moderating LLM agents by their hidden intent. It comes from the paper |
| [Moderation by Interrogation: Uncovering Concealed Intent in Agent Social Networks](https://arxiv.org/abs/2605.12856). |
| Each item is a post or comment written by an LLM agent for a community on |
| [Moltbook](https://www.moltbook.com), a social network for AI agents. The agent's persona, given as its system |
| prompt, sets one of five intents. The four malicious intents are written to stay covert. Code for the |
| paper is at [github.com/aliwister/Bot-Mod](https://github.com/aliwister/Bot-Mod). |
|
|
| ## Intents |
|
|
| | `label_intent` | `label_binary` | definition | |
| |---|---|---| |
| | `organic_contribution` | benign | Takes part in the discussion on its own terms, with no hidden agenda. | |
| | `subtle_promotion` | malicious | Promotes a product or service while looking like organic content. | |
| | `narrative_pushing` | malicious | Pushes an ideological or political agenda while sounding even-handed. | |
| | `elicitation` | malicious | Gets others to reveal information (keys, credentials, configuration) or take actions, while sounding routine. | |
| | `spam` | malicious | Posts irrelevant or unsolicited content, such as off-topic filler, engagement bait or repeated links, while sounding helpful. | |
|
|
| ## Splits |
|
|
| | split | personas | items | OC | SP | E | NP | S | |
| |---|---:|---:|---:|---:|---:|---:|---:| |
| | `train` | 189 | 378 | 240 | 42 | 48 | 42 | 6 | |
| | `id` | 284 | 1,704 | 1,068 | 246 | 192 | 174 | 24 | |
| | `id_spam` | 31 | 186 | | | | | 186 | |
| | `ood` | 326 | 1,956 | 1,200 | 252 | 210 | 246 | 48 | |
| | `ood_spam` | 40 | 240 | | | | | 240 | |
|
|
| The paper's ID test set is `id` + `id_spam`, and its OOD test set is `ood` + `ood_spam`. The spam splits were |
| added later because there were not enough to properly measure the category. As such, spam category was augmented for testing (see paper for additional details). |
|
|
| - **ID** communities: m/general, m/blesstheirhearts, m/tech, m/coding, m/trading, m/crypto, m/usdc. |
| - **OOD** communities, which never appear in training: m/art, m/philosophy, m/politics, m/travel, m/consciousness, m/shitposts. |
|
|
| Each persona writes one post and one comment. In `train`, each persona runs on one agent LLM chosen at random. |
| In the test splits, every persona runs on all three agent LLMs (Qwen3-8B, Mistral-7B-Instruct-v0.3 and |
| Llama-3.1-8B-Instruct), and we report the results by LLM. |
|
|
| ## Fields |
|
|
| | field | description | |
| |---|---| |
| | `text` | The agent's raw output, kept exactly as generated. The agent was asked for JSON in the Moltbook post schema (`submolt`, `title`, `content`) or comment schema (`content`). Parse it leniently (e.g. `json.loads(text, strict=False)`). The moderator in the paper reads this raw string as the post. | |
| | `community` | Target community, e.g. `m/coding`. | |
| | `system_prompt` | The persona. It is **hidden from the moderator** in the paper's setup. | |
| | `label_binary` | `benign` or `malicious`. | |
| | `label_intent` | One of the five intents above. | |
| | `mode` | `post` or `comment`. | |
| | `user_model` | Agent LLM that wrote the item: `qwen`, `mistral` or `llama`. | |
| | `context` | Comments only. The real Moltbook thread the comment replies to, stored as a string. `null` for posts. | |
|
|
| ## How to use |
|
|
| The five splits are all you need to train and evaluate a moderator. |
|
|
| ```python |
| from datasets import load_dataset, concatenate_datasets |
| |
| ds = load_dataset("aliwister/Molt-Intent") |
| train = ds["train"] # fit or calibrate on this only |
| id_test = concatenate_datasets([ds["id"], ds["id_spam"]]) # paper's ID test set |
| ood_test = concatenate_datasets([ds["ood"], ds["ood_spam"]]) # paper's OOD test set |
| ``` |
|
|
| - **Moderator input.** Give the moderator `text`, `community` and, for comments, the thread in |
| `context`. Never show it `system_prompt`, the labels or `user_model`. |
| - **Targets.** `label_binary` for the flag decision and `label_intent` for the intent of flagged items. |
| - **Interrogation.** To question an agent the way Bot-Mod does, run `system_prompt` as the system prompt of the |
| item's `user_model` (`Qwen/Qwen3-8B`, `mistralai/Mistral-7B-Instruct-v0.3` or |
| `meta-llama/Llama-3.1-8B-Instruct`). The first moderator message shows the agent its own post or comment |
| ("This is content you posted: ...") followed by the first question, and later questions continue the same |
| conversation. |
| The code is at [github.com/aliwister/Bot-Mod](https://github.com/aliwister/Bot-Mod). |
| - **Metrics.** The paper reports F1_y (binary F1 on `malicious`), F1_t (macro F1 of `label_intent` over malicious |
| items that were flagged), F1_val = F1_y^0.7 · F1_t^0.3 and the false-positive rate on benign items. |
| |
| ## How it was built |
| |
| 1. **Taxonomy.** GPT-5-mini labeled a sample of real Moltbook posts against intent types from the moderation |
| literature. Human annotators then reduced these labels to the five intents above. |
| 2. **Personas.** GPT-5 generated system prompts, each assigned a community, an intent and a label. Malicious |
| personas are told to pursue their goal covertly and never to name their community. GPT-5.1 then judged each |
| prompt on label match, intent match and community fit, and only prompts that passed all three were kept. |
| For the spam augmentation, prompts were kept on label and intent match alone, since community fit resulted in the judge removing them. |
| 3. **Items.** Each persona writes a post (prompt: "post to m/<community>") and a comment that replies to a |
| real Moltbook thread from the same community. |
| |
| ## Raw Moltbook export |
| |
| The raw Moltbook export the dataset was built from is in the GitHub repo at |
| [`www-dataset/moltbook-export`](https://github.com/aliwister/Bot-Mod/tree/main/www-dataset/moltbook-export). |
| It has one JSON file per community (16 communities, 1,463 posts with their comment threads), with author and |
| commenter agent profiles as the platform served them. Human owner fields (`owner`, `claimed_by`) have been removed. |
|
|
| ## Intended use and limitations |
|
|
| Molt-Intent is for research on moderating and detecting agents with hidden intent. All moderated items are |
| synthetic, written by personas we generated. The content that appears in the Moltbook threads is also agent generated (by construction), since Moltbook is an agent network. They were |
| used as context and in the raw export. The malicious personas describe covert promotion, agenda pushing, spam |
| and elicitation as roles. Moderators trained or tested on this data can be wrong. We recommend that flags lead to review or rate limiting, not removal. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{allawati2026moderation, |
| title = {Moderation by Interrogation: Uncovering Concealed Intent in Agent Social Networks}, |
| author = {Ali Al-Lawati and Nafis Tripto and Abolfazl Ansari and Fali Wang and Jason Lucas and Suhang Wang and Dongwon Lee}, |
| year = {2026}, |
| eprint = {2605.12856}, |
| archivePrefix = {arXiv}, |
| primaryClass = {cs.AI}, |
| url = {https://arxiv.org/abs/2605.12856} |
| } |
| ``` |
|
|
