--- pretty_name: Mental Models for Multi-Agent Systems language: - en - zh license: other license_name: mixed-cc-by-4.0-mit-apache-2.0 task_categories: - text-generation - question-answering - visual-question-answering tags: - theory-of-mind - multi-agent - mental-modeling - social-reasoning - role-playing configs: - config_name: sotopia-turn-rewards default: true data_files: - split: train path: data/sotopia/sotopia_turn_rewards_v3.jsonl - config_name: sotopia-mental-personas data_files: - split: train path: data/sotopia/mental_model_persona_dataset.jsonl - config_name: bigtom data_files: - split: train path: data/bigtom/bigtom_qwen_5k_annotated.jsonl - config_name: mmrole-belief data_files: - split: train path: data/mmrole/train/belief_prediction.jsonl - split: validation path: data/mmrole/val/belief_prediction.jsonl - split: test_in path: data/mmrole/test_in/belief_prediction.jsonl - split: test_out path: data/mmrole/test_out/belief_prediction.jsonl - split: test_official path: data/mmrole/official_test/belief_prediction.jsonl - config_name: mmrole-preference data_files: - split: train path: data/mmrole/train/preference_pairs.jsonl - split: validation path: data/mmrole/val/preference_pairs.jsonl - split: test_in path: data/mmrole/test_in/preference_pairs.jsonl - split: test_out path: data/mmrole/test_out/preference_pairs.jsonl - split: test_official path: data/mmrole/official_test/preference_pairs.jsonl - config_name: mmrole-salience data_files: - split: train path: data/mmrole/train/salience_prediction.jsonl - split: validation path: data/mmrole/val/salience_prediction.jsonl - split: test_in path: data/mmrole/test_in/salience_prediction.jsonl - split: test_out path: data/mmrole/test_out/salience_prediction.jsonl - split: test_official path: data/mmrole/official_test/salience_prediction.jsonl - config_name: mmrole-probe-qa data_files: - split: train path: data/mmrole/train/probe_qa.jsonl - split: validation path: data/mmrole/val/probe_qa.jsonl - split: test_in path: data/mmrole/test_in/probe_qa.jsonl - split: test_out path: data/mmrole/test_out/probe_qa.jsonl - split: test_official path: data/mmrole/official_test/probe_qa.jsonl - config_name: mmrole-raw-annotations data_files: - split: train path: data/mmrole/train/raw_annotated.jsonl - split: validation path: data/mmrole/val/raw_annotated.jsonl - split: test_in path: data/mmrole/test_in/raw_annotated.jsonl - split: test_out path: data/mmrole/test_out/raw_annotated.jsonl - config_name: mmrole-raw-official data_files: - split: test path: data/mmrole/official_test/raw_annotated.jsonl - config_name: mmrole-reward-belief data_files: - split: train path: data/mmrole/train/belief_prediction_mmrole_reward_openai.jsonl - split: validation path: data/mmrole/val/belief_prediction_mmrole_reward_openai.jsonl - split: test_in path: data/mmrole/test_in/belief_prediction_mmrole_reward_openai.jsonl - config_name: mmrole-reward-preference data_files: - split: train path: data/mmrole/train/preference_pairs_mmrole_reward_openai.jsonl - split: validation path: data/mmrole/val/preference_pairs_mmrole_reward_openai.jsonl - split: test_in path: data/mmrole/test_in/preference_pairs_mmrole_reward_openai.jsonl --- # Mental Model Annotation Dataset ## Dataset for *Mental Models for Multi-Agent Systems* This is the official annotation release accompanying the NeurIPS 2026 paper ***Mental Models for Multi-Agent Systems***. **Paper resources:** [Project page](https://hananshafi.github.io/Mental-Models/) | [Code](https://github.com/hananshafi/Mental-Models) | Paper and arXiv links will be added upon release. The paper studies explicit, recursive mental representations for multi-agent decision-making. This dataset contains the mental-state, reward, rationale, and preference supervision used for its SOTOPIA, BigToM, and MMRole experiments. The files augment these established benchmarks; they are not a new replacement for the original benchmark datasets. Independent configurations preserve each benchmark's schema and remain compatible with the Hugging Face Dataset Viewer. ## Contents | Benchmark | Configuration | Contents | |---|---|---| | SOTOPIA | `sotopia-turn-rewards` | 1,647 interaction episodes with 16,166 turn-level reward records and mental-state supervision | | SOTOPIA | `sotopia-mental-personas` | 500 mental-model persona examples | | BigToM | `bigtom` | 9,964 annotated conditions from 4,982 paired scenarios | | MMRole | `mmrole-belief` | First- and second-order belief targets | | MMRole | `mmrole-preference` | Preferred responses and hard negatives | | MMRole | `mmrole-salience` | Visual perspective and salience targets | | MMRole | `mmrole-probe-qa` | Theory-of-Mind diagnostic questions | | MMRole | `mmrole-raw-annotations` | Full validated mental-state annotations | | MMRole | `mmrole-raw-official` | Official test annotations and test metadata | | MMRole | `mmrole-reward-belief` | Belief examples with eight-dimensional reward labels | | MMRole | `mmrole-reward-preference` | Preference pairs with eight-dimensional reward labels | ## Loading ```python from datasets import load_dataset sotopia = load_dataset( "hanangani/Mental-Model-Annotation-Dataset", "sotopia-turn-rewards", ) bigtom = load_dataset("hanangani/Mental-Model-Annotation-Dataset", "bigtom") mmrole = load_dataset("hanangani/Mental-Model-Annotation-Dataset", "mmrole-belief") ``` Download the original files without schema conversion when using the released training scripts: ```bash hf download hanangani/Mental-Model-Annotation-Dataset \ --repo-type dataset \ --local-dir Mental-Models-data ``` ## MMRole images This repository does not duplicate MMRole's 11,032 source images. MMRole rows retain the upstream `image` and `image_local` references. Download the images from [`YanqiDai/MMRole_dataset`](https://huggingface.co/datasets/YanqiDai/MMRole_dataset) and follow its instructions for any referenced MS-COCO files. ## Validation All released JSONL records were parsed before upload. `MANIFEST.json` records the row count, byte size, and SHA-256 checksum of every staged file. Temporary files, failed annotation attempts, smoke tests, model checkpoints, and API credentials are excluded. ## Sources and licenses This is a derived, mixed-source research dataset. See [`DATA_SOURCES.md`](DATA_SOURCES.md) for provenance, upstream licenses, and redistribution notes. No additional rights are granted for upstream content. ## Citation ```bibtex @inproceedings{gani2026mentalmodels, title = {Mental Models for Multi-Agent Systems}, author = {Gani, Hanan and Shao, Lulu and Chandraker, Manmohan}, booktitle = {Advances in Neural Information Processing Systems}, year = {2026} } ```