Semiotic Reflexive Transformer (SRT) - Stage 1 Data

SRT Stage 1 Validation Data: Description and Documentation Overview This data package contains the synthetic datasets, trained model checkpoints, and validation results for Stage 1 of the Semiotic-Reflexive Transformer (SRT) project. The SRT is a neural architecture that embeds Peircean semiotic decomposition, metapragmatic divergence tracking, and catastrophe-theoretic bifurcation estimation into the transformer computational graph. These datasets were designed to validate that each architectural module learns its intended semiotic function under controlled conditions with known ground-truth signals. All data was generated synthetically with planted divergence signals to enable precise, falsifiable evaluation of four core architectural claims: (1) that the Semiotic Embedding Layer produces distinct, interpretable subspaces corresponding to Peircean categories, (2) that community-conditioned interpretants differentiate contested from neutral terms, (3) that the Metapragmatic Attention Head tracks accumulating meaning divergence across token positions, and (4) that the Bifurcation Estimation Network detects sharp transitions between single-interpretation and dual-interpretation regimes. Dataset Descriptions Dataset A: Binary Community Lexicon Purpose: Tests subspace specialization (Claim 1) and community differentiation (Claim 2). Structure: Two synthetic interpretive communities share a 200-word vocabulary. Twenty words are designated "contested" — their co-occurrence statistics differ systematically between communities. Contested words co-occur with positive-valence terms in Community 0 and negative-valence terms in Community 1. The remaining 79 non-reserved words are "neutral" with identical distributions across communities. Ground truth: Contested words carry rtrue = 0.8; neutral words carry rtrue = 0.0. Samples: 5,000 sequences (4,500 train / 500 validation). Format: JSONL. Each record contains token_ids, community_id, chain_labels, chain_divergence, attractor_labels, and r_true. Dataset B: Gradual Divergence Ramp Purpose: Tests divergence tracking (Claim 3). Structure: Each sequence contains 128 tokens. Divergence increases monotonically across three phases: positions 0–32 are low-divergence (rtrue ≈ 0.0), positions 32–64 show mild divergence (rtrue ramps linearly from 0.0 to 0.5), and positions 64–128 show strong divergence (rtrue ramps from 0.5 to 1.0). The ramp is implemented by interpolating between shared-vocabulary tokens (low divergence) and community-specific tokens (high divergence) as the sequence progresses. Ground truth: Per-position rtrue values following the piecewise linear ramp. Samples: 5,000 sequences (4,500 train / 500 validation). Format: JSONL. Each record contains token_ids, community_id, chain_divergence (per-position float following the ramp), and source identifier. Dataset C: Bifurcation Events Purpose: Tests bifurcation detection (Claim 4). Structure: Each sequence contains a single sharp bifurcation point at a randomly chosen position k. Before position k, all tokens are drawn from a shared vocabulary with uniform co-occurrence statistics (rtrue ≈ 0.0). At position k, a trigger token is inserted, and all subsequent tokens are drawn from community-specific distributions (rtrue ≈ 0.7). Ground truth: Per-position rtrue values (≈ 0.0 before k, ≈ 0.7 after k), plus the bifurcation position k stored in the metadata field. Samples: 5,000 sequences (4,500 train / 500 validation). Format: JSONL. Each record contains token_ids, community_id, chain_divergence, attractor_labels, r_true, and metadata (including bifurcation_position). Model Checkpoint Architecture: SRT at TINY preset (31.6M parameters). Configuration: dmodel = 512, 6 layers, 8 attention heads, dsub = 128, 16 community embeddings, SwiGLU activations, RMSNorm, RoPE positional encoding. Training: 20 epochs on all three datasets jointly, AdamW optimizer with cosine learning rate scheduling, composite loss function (cross-entropy + chain consistency + attractor basin + bifurcation estimation; iconic grounding disabled). Hardware: Apple M-series GPU via MPS backend (~2 hours training time). Format: PyTorch .pt checkpoint files containing model state dict, optimizer state, scheduler state, and training metadata. Validation Results The stage1_results.json file contains the complete quantitative results from Stage 1 validation: Test Metric Result Threshold Status 1.3.1 Subspace Specialization Linear probing margin (min across 4 tasks) 0.155 ≥ 0.15 PASS 1.3.2 Community Differentiation Contested/neutral cosine distance ratio 3.28× ≥ 3.0× PASS 1.3.3 Divergence Tracking Spearman ρ with ground-truth ramp 0.822 ≥ 0.6 PASS 1.3.4 Bifurcation Detection Regime classification accuracy 100.0% ≥ 75% PASS 1.3.4 Bifurcation Detection Mean r̂ difference (post − pre) 0.659 > 0.2 PASS Reproduction To regenerate the datasets from scratch: python scripts/generate_synthetic.py --output data/synthetic --seed 42 Random seed: All data generation uses seed 42 by default for full reproducibility. File Manifest srt_stage1_data_v1.0/ ├── data/synthetic/ │ ├── train/ │ │ ├── dataset_a.jsonl (Binary community lexicon, 4,500 samples) │ │ ├── dataset_b.jsonl (Gradual divergence ramp, 4,500 samples) │ │ ├── dataset_c.jsonl (Bifurcation events, 4,500 samples) │ │ └── combined.jsonl (All training data combined, 13,500 samples) │ └── val/ │ ├── dataset_a.jsonl (500 samples) │ ├── dataset_b.jsonl (500 samples) │ ├── dataset_c.jsonl (500 samples) │ └── combined.jsonl (All validation data combined, 1,500 samples) ├── results/ │ └── stage1_results.json (Complete validation metrics) ├── model_config/ │ └── config.json (TINY preset architecture hyperparameters) ├── scripts/ │ └── generate_synthetic.py (Deterministic dataset generation script) └── DATASHEET.md (Full datasheet per Gebru et al., 2021) Citation If you use this data or the SRT architecture in your research, please cite: Lancaster, J. B. (2026). The Semiotic-Reflexive Transformer: A Neural Architecture for Detecting and Modulating Meaning Divergence Across Interpretive Communities. SSRN Electronic Journal. Repository https://github.com/space-bacon/Semiotic-Reflexive-Transformer Contact Burton Lancaster — [e-mail] 第一阶段验证数据集：说明与文档 ## 概述 本数据包包含Semiotic-Reflexive Transformer（SRT）项目第一阶段的合成数据集、训练模型 checkpoint 与验证结果。SRT是一种将皮尔斯符号学分解、元语用分歧追踪与灾变理论分岔估计嵌入Transformer（Transformer）计算图的神经架构。本系列数据集旨在验证：在带有已知真值信号的受控条件下，每个架构模块能否学会其预设的符号学功能。 所有数据均为带有预设分歧信号的合成数据，用于对四项核心架构主张进行精准、可证伪的评估：(1) 符号嵌入层可生成对应皮尔斯范畴的可区分、可解释子空间；(2) 社群条件化解释项可区分争议性术语与中性术语；(3) 元语用注意力头可追踪不同Token（Token）位置上累积的语义分歧；(4) 分岔估计网络可检测单解释态与双解释态之间的急剧转变。 ## 数据集说明 ### 数据集A：二元社群词典 **用途**：测试子空间专业化（主张1）与社群分化（主张2）。 **结构**：两个合成解释社群共享一份包含200个词汇的词典。其中20个词汇被标记为“争议性词汇”——它们在两个社群中的共现统计存在系统性差异：社群0中争议性词汇与积极情感词汇共现，社群1中则与消极情感词汇共现。剩余79个非预留词汇为“中性词汇”，在两个社群中的分布完全一致。 **真值标签**：争议性词汇的真实分歧值$r_{ ext{true}}=0.8$；中性词汇的真实分歧值$r_{ ext{true}}=0.0$。 **样本量**：5000条序列（4500条训练/500条验证）。 **格式**：JSONL。每条记录包含`token_ids`、`community_id`、`chain_labels`、`chain_divergence`、`attractor_labels`与`r_true`字段。 ### 数据集B：渐进式分歧斜坡 **用途**：测试分歧追踪能力（主张3）。 **结构**：每条序列包含128个Token。分歧随序列推进分为三个单调递增的阶段：位置0-32为低分歧阶段（$r_{ ext{true}}≈0.0$）；位置32-64为轻微分歧阶段（$r_{ ext{true}}$从0.0线性上升至0.5）；位置64-128为高分歧阶段（$r_{ ext{true}}$从0.5上升至1.0）。该斜坡通过在共享词汇Token与社群专属Token之间插值实现。 **真值标签**：每个位置的$r_{ ext{true}}$值遵循分段线性斜坡分布。 **样本量**：5000条序列（4500条训练/500条验证）。 **格式**：JSONL。每条记录包含`token_ids`、`community_id`、`chain_divergence`（每个位置的浮点型分歧值，遵循斜坡分布）与`source identifier`字段。 ### 数据集C：分岔事件 **用途**：测试分岔检测能力（主张4）。 **结构**：每条序列在随机选定的位置$k$处存在一个单一的急剧分岔点。在位置$k$之前，所有Token均从共现统计一致的共享词汇中抽取（$r_{ ext{true}}≈0.0$）；在位置$k$处插入触发Token，后续所有Token均从社群专属分布中抽取（$r_{ ext{true}}≈0.7$）。 **真值标签**：每个位置的$r_{ ext{true}}$值（位置$k$前约为0.0，位置$k$后约为0.7），以及存储于元数据字段中的分岔位置$k$。 **样本量**：5000条序列（4500条训练/500条验证）。 **格式**：JSONL。每条记录包含`token_ids`、`community_id`、`chain_divergence`、`attractor_labels`、`r_true`与`metadata`（包含`bifurcation_position`字段）。 ## 模型检查点 **架构**：TINY预设参数的SRT模型（3160万参数）。 **配置**：$d_{ ext{model}}=512$，6层编码器，8个注意力头，$d_{ ext{sub}}=128$，16个社群嵌入，SwiGLU（SwiGLU）激活函数，RMSNorm（RMSNorm）归一化，RoPE（RoPE）位置编码。 **训练设置**：在全部三个数据集上联合训练20个epoch（epoch），使用AdamW（AdamW）优化器与余弦学习率调度，复合损失函数（交叉熵+序列一致性+吸引盆（attractor basin）+分岔估计；禁用具象接地（iconic grounding））。 **硬件**：基于MPS后端的Apple M系列GPU（训练耗时约2小时）。 **格式**：PyTorch .pt格式检查点文件，包含模型状态字典、优化器状态、调度器状态与训练元数据。 ## 验证结果 `stage1_results.json`文件包含第一阶段验证的全部定量结果： | 测试项 | 指标 | 结果 | 阈值 | 状态 | | --- | --- | --- | --- | --- | | 1.3.1 子空间专业化 | 线性探测边际（4项任务最小值） | 0.155 | ≥0.15 | PASS | | 1.3.2 社群分化 | 争议/中性词汇余弦距离比 | 3.28× | ≥3.0× | PASS | | 1.3.3 分歧追踪 | 与真实斜坡的斯皮尔曼$ ho$相关系数 | 0.822 | ≥0.6 | PASS | | 1.3.4 分岔检测 | 状态分类准确率 | 100.0% | ≥75% | PASS | | 1.3.4 分岔检测 | 平均$hat{r}$差值（后段 - 前段） | 0.659 | >0.2 | PASS | ## 复现方法 若需从头重新生成数据集，可执行以下命令： bash python scripts/generate_synthetic.py --output data/synthetic --seed 42 **随机种子**：所有数据生成过程默认使用种子42，以保证完全可复现。 ## 文件清单 srt_stage1_data_v1.0/ ├── data/synthetic/ │ ├── train/ │ │ ├── dataset_a.jsonl # 二元社群词典，4500条样本 │ │ ├── dataset_b.jsonl # 渐进式分歧斜坡，4500条样本 │ │ ├── dataset_c.jsonl # 分岔事件，4500条样本 │ │ └── combined.jsonl # 全部训练数据合并，13500条样本 │ └── val/ │ ├── dataset_a.jsonl # 500条样本 │ ├── dataset_b.jsonl # 500条样本 │ ├── dataset_c.jsonl # 500条样本 │ └── combined.jsonl # 全部验证数据合并，1500条样本 ├── results/ │ └── stage1_results.json # 完整验证指标 ├── model_config/ │ └── config.json # TINY预设架构超参数 ├── scripts/ │ └── generate_synthetic.py # 确定性数据集生成脚本 └── DATASHEET.md # 符合Gebru等人2021年标准的完整数据集说明书 ## 引用 若您在研究中使用本数据集或SRT架构，请引用以下文献： > Lancaster, J. B. (2026). The Semiotic-Reflexive Transformer: A Neural Architecture for Detecting and Modulating Meaning Divergence Across Interpretive Communities. SSRN Electronic Journal. ## 代码仓库 https://github.com/space-bacon/Semiotic-Reflexive-Transformer ## 联系方式 Burton Lancaster — [e-mail]