Genesis-AI-Labs/tinyvm-tier1

Tiny-VM Tier 1 (Register Traces) 是一个合成数据集，包含直线型Tiny-VM程序的完整执行跟踪和预渲染训练提示。该数据集属于FANCLatent State as Computer实验课程的第一层，专注于在有界程序长度下跟踪寄存器文件。数据集包括200,000个训练程序和140,000个分层评估程序（按程序长度分7个桶，每个桶20,000个，程序长度n ∈ {8, 16, 32, 48, 64, 96, 128}）。生成器使用tinyvm.generators.gen_register_trace，支持操作如LOAD/ADD/SUB/MOV/PRINT，无控制流。每行数据包含完整程序（指令列表）、执行跟踪（每步寄存器快照、输出流、停止标志）以及直接模式预渲染提示（令牌ID和表面文本）。数据集具有确定性，每行数据可通过元种子和轴参数重建，并具有字节完整性，通过SHA-256哈希验证。数据集适用于文本生成和推理任务，旨在支持程序执行和可解释性研究。Tiny-VM Tier 1 — Register Traces is a synthetic dataset of straight-line Tiny-VM programs with full execution traces and pre-rendered training prompts. Tier 1 of the FANC Latent State as Computer experimental curriculum, focused on register-file tracking under bounded program length. It includes 200,000 train programs and 140,000 stratified eval programs (7 buckets × 20,000, one per program length n ∈ {8, 16, 32, 48, 64, 96, 128}). The generator uses tinyvm.generators.gen_register_trace (supporting LOAD / ADD / SUB / MOV / PRINT, no control flow). Each row contains a full Program (instruction list), ExecutionTrace (per-step register snapshots, output stream, halted flag), and direct-mode pre-rendered prompt (token IDs and surface text). The dataset is deterministic, with each row reconstructable from meta.seed and meta.axes, and ensures byte-integrity via SHA-256 hashes. It is designed for text-generation and reasoning tasks, supporting research in program execution and interpretability.