0xRyanxx/agent_consensus_stacked

该数据集名为Agent Consensus — Pre-stacked Hidden States，是一个预堆叠的Qwen3-8B隐藏状态激活数据集，专门用于多智能体活性探测扫描，支持论文Why Do LLM Agents Stall?中的机制研究，该研究关注标量拜占庭共识和GSM8K数学QA共识中的公共理性与潜在状态中介。数据集包含多个文件：隐藏状态文件（如h_pre_layer{0,6,12,18,24,30,35}.npy和h_pre_masked_layer{...}.npy，形状约为(3.85M, 4096)，数据类型为float16），这些文件存储了在理性生成前最后一个提示词处的隐藏状态，支持不同变压器层；清单文件（如*.paths.txt）提供加载器映射信息，将.npy文件中的行索引与对应的智能体×轮次NPZ文件关联；辅助文件包括探测数据集（probe_dataset.parquet）、预计算的文本基线（text_baselines_scalar.pkl和text_baselines_gsm8k.pkl）以及GSM8K测试子集（gsm8k_subset.json）。数据集旨在简化探测训练过程，通过预堆叠隐藏状态减少文件打开延迟，适用于机器学习研究，特别是自然语言处理和机制可解释性领域。The dataset is named Agent Consensus — Pre-stacked Hidden States, which consists of pre-stacked Qwen3-8B hidden-state activations for multi-agent liveness probe sweeps, supporting the mechanistic study in the paper Why Do LLM Agents Stall? that investigates public-rationale-versus-latent-state mediation in scalar Byzantine consensus and GSM8K math QA consensus. It includes files such as hidden state files (e.g., h_pre_layer{0,6,12,18,24,30,35}.npy and h_pre_masked_layer{...}.npy, with shapes around (3.85M, 4096) and dtype float16), which store hidden states at the last prompt token before rationale generation across different transformer layers; manifest files (e.g., *.paths.txt) that provide mapping information for loaders to associate row indices in .npy files with corresponding agent×round NPZ files; and auxiliary files like the probe dataset (probe_dataset.parquet), precomputed text baselines (text_baselines_scalar.pkl and text_baselines_gsm8k.pkl), and a GSM8K test subset (gsm8k_subset.json). The dataset aims to streamline probe training by reducing file-open latency through pre-stacked hidden states, making it suitable for machine learning research, particularly in natural language processing and mechanistic interpretability.