danielfein/WildDelusion
收藏资源简介:
--- license: mit task_categories: - text-classification tags: - mental-health - delusion - wildchat - probing --- # WildDelusion High-scoring messages and conversations from [WildChat-4.8M-Full](https://huggingface.co/datasets/yuntian-deng/WildChat-4.8M-Full) flagged by a diff-mean linear probe trained to detect user-endorsed delusional content. ## Probe details ```json { "probe_type": "diffmean_wc_length_matched_negatives", "probe_dataset": "danielfein/MHProbes", "probe_config": "delusion_adversarial_pairs_gpt54mini_subtle_v1", "model_id": "meta-llama/Llama-3.2-1B-Instruct", "target_layer_1based": 13, "score_threshold": 9.5, "wildchat_rows_scanned": "~1M / 4.74M (21%)" } ``` ## Configs ### `messages` One row per flagged user message. Columns: `score`, `text`, `message_idx`, `conversation_hash`, `model`, `row_idx`, `conversation` (full JSON), `language`, `turn`, `country`. ### `conversations` One row per unique conversation containing ≥1 flagged turn. Columns: `conversation_hash`, `row_idx`, `model`, `language`, `turn`, `country`, `conversation` (full JSON), `flagged_turns` (JSON list), `max_score`, `n_flagged_turns`. ## Note This is a partial scan (~12% of WildChat). Dataset will be updated as the scan completes.




