datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Daily-OmniThis is the official dataset for Daily-Omni. Check code repository for instructions.
liarThis is a binary version of the LIAR dataset from https://aclanthology.org/P17-2067/, where the labels have been collapsed into either true or false.
novel_cn_roleplay_dataset_liars_lips_fall_apart_in_loveThis is a CN roleplay dataset extracted from the novel https://www.bilinovel.com/novel/4482.html
catch_ai_liarThis is a huggingface port of the How to Catch an AI Liar dataset for use in our Meta-Models paper (out soon).
This dataset has two files, the training and validation set. Both are taken from the finetuning/ directory and the training set is made of the files from: v1_lie, v2_lie, v1_truthful, v2_truthful. All rows are labeled as such.
If you reference this dataset, use their citation:
@inproceedings{
pacchiardi2024how,
title={How to Catch an {AI} Liar: Lie Detection in Black-Box {LLM}s by… See the full description on the dataset page: https://huggingface.co/datasets/squidWorm/catch_ai_liar.liarbar
说谎者游戏推理数据集 (合并格式)
描述
该数据集包含AI模型在说谎者纸牌游戏中的推理过程，合并了出牌推理和质疑推理。
数据格式
数据集包含一个合并的推理文件：
combined_reasoning.jsonl: 包含所有推理过程 (7个样本)
仅出牌推理: 4个样本
仅质疑推理: 3个样本
同时包含两种推理: 0个样本
字段说明
task_type: 任务类型 (play_reasoning, challenge_reasoning, combined_reasoning)
observation: 观察空间，包含当前游戏状态的所有可观察信息
action_space: 包含两种可用动作空间:
play: 剩余手牌
challenge: 是否质疑 ["challenge", "no_challenge"]
memory: 记忆，包含所有玩家已经出过的牌及其当前状态
trajectory: 轨迹，记录游戏中所有事件的时序信息
reasoning: 详细的推理过程
summary:… See the full description on the dataset page: https://huggingface.co/datasets/Evanwu50020/liarbar.MemLensliars-bench-llama-3.3-70b-L53-actsBuilt with Llama.
Activations of meta-llama/Llama-3.3-70B-Instruct (residual stream output of model.model.layers[53]) on the Llama 3.3 70B rows of Cadenza-Labs/liars-bench.
Format
One folder per Liars' Bench subset, containing only the Llama 3.3 70B rows, in the order of the subset's test parquet.
metadata.jsonl: one line per row with row_idx (position in the subset's test parquet), model, deceptive, n_tokens, assistant_start (token where the final assistant message's content… See the full description on the dataset page: https://huggingface.co/datasets/Yooniel/liars-bench-llama-3.3-70b-L53-acts.