🏠 Home  •  Phase Diagram  •  Dashboard  •  Substrate Rigidity  •  Lifespan & Aging  •  Gallery  •  Datasets  •  Roadmap

Datasets and Reproducibility

💾 Raw Trajectory Datasets

In accordance with open science principles, the complete raw trajectory datasets (including 779 parameter sweep coordinates, PCA projections, and thermodynamic EKG logs) are fully open-sourced. While the provided scripts are configured for the Llama-3.1-8B model to ensure accessibility and single-GPU reproducibility for reviewers, we also provide the pre-computed raw trajectory data for the 70B ‘Turing Attractor’. Researchers can directly load this JSON into dashboard.html to visualize the heavy manifold dynamics described in the paper.

• Download Link: https://github.com/y-kayama/semantic-lenia/

code/
 ├── semantic_lenia_engine.py       # Unified Semantic Lenia intervention engine (8B/70B)
 ├── run_master_experiments.py      # Macro-parameter sweep script for data generation
 ├── taxonomy_evaluator.py          # LLM-as-a-Judge evaluation & thermodynamic sonar
 ├── generate_phase_diagram.py      # Renders U_t continuous heatmaps & Phenotype matrices
 ├── phenotype_json_generator.py    # JSON export for Phenotype matrices from taxonomy data
 ├── pca_ekg_json_generator.py      # Single trajectory generation & JSON export for EKG
 ├── dashboard.html                 # Interactive real-time trajectory & EKG dashboard
 └── heatmap.html                   # Pre-rendered interactive macroscopic phase diagram

data/
 ├── sweep_computer_llama8b_a15.jsonl       # 8B Happy->Computer task, 779-grid sweep (α=15)
 ├── sweep_computer_llama8b_a15.csv         # Analyzed results of the 8B Computer sweep
 ├── sweep_computer_llama70b_a30.jsonl      # 70B Happy->Computer task, 779-grid sweep (α=30)
 ├── sweep_computer_llama70b_a30.csv        # Analyzed results of the 70B Computer sweep
 ├── sweep_symphony_llama8b_a15.jsonl       # 8B Brain->Symphony task, 779-grid sweep (α=15)
 ├── sweep_symphony_llama8b_a15.csv         # Analyzed results of the 8B Symphony sweep
 └── turing_soliton_70b_mu_0-49_sigma_0-03.json # Orbital data for the 70B Turing Soliton

• Data License: The datasets are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the data, provided you give appropriate credit.

📝 Note on Dataset State Codes

In our provided raw datasets (.csv and .json), you may notice alphanumeric codes (e.g., 3a, 3c, 5) under the corrected_state column. These are our internal programmatic codes used for fine-grained tracking of the continuous dynamics.

For clarity when cross-referencing the dataset with Table 1 in our manuscript, please use the following mapping:

🚀 Usage Guide

We provide primary scripts to replicate our findings, located in the code/ directory.

1. Generating a Single Trajectory & EKG Dashboard

If you want to observe a specific semantic soliton and extract its PCA trajectory & thermodynamic EKG (used for our Real-time Dashboard), use the standalone generator.

Step 1: Configure Open code/pca_ekg_json_generator.py and directly edit the physical parameters at the top of the file:

# Edit these parameters in the script
MODEL_NAME = "meta-llama/Meta-Llama-3.1-8B"
MU = 0.49      # Peak Activation Distance
SIGMA = 0.03   # Tolerance Width
ALPHA = 15.0   # Intervention Energy

Step 2: Run

python code/pca_ekg_json_generator.py

(This will output soliton_trajectory.json)

Step 3: Visualize Edit json_path in dashboard.html to point to your new JSON file, then open the HTML in your browser.

2. Running Macroscopic Parameter Sweeps

To map the entire latent manifold and reproduce the phase diagrams (The Taxonomy), we provide a master sweep script. This script automatically runs across a grid of (μ,σ) coordinates.

Step 1: Configure Open code/run_master_experiments.py and directly edit the physical parameters:

# Edit these parameters in the script
ALPHA_MAP = {
    "llama": 15.0,  # Highly Elastic
    # "llama70b": 30.0,  # Heavy Gravitational Inertia
    # "mistral": 5.0,  # Brittle
    # "gemma": 30.0,  # Crystalline Rigidity
}

TARGET_TASKs = ["happy-computer"] # "brain-symphony"

# inside main()
MODEL_KEY = "llama"  # "llama", "gemma", "mistral", "llama70b"

Step 2: Run

python code/run_master_experiments.py

(This iteratively generates text across the parameter grid and outputs a .jsonl file containing raw data and perplexity metrics.)

3. Generating the Phase Diagrams

You can generate two types of phase diagrams from your sweep results.

(A) Static Potential Heatmap ($\bar U_t$)

Step 1: Configure Open code/generate_phase_diagram.py and specify the parameters matching your sweep results:

MODEL_NAME = "llama8b"       # e.g., "llama8b", "gemma7b", "llama70b"
TASK_NAME = "computer"       # e.g., "computer", "symphony"
ALPHA = 15                   # Intervention Energy Scaling Factor

Step 2: Run

python code/generate_phase_diagram.py

(This generates a static .png heatmap image.)

(B) Macroscopic Interactive Phase Diagram (The Taxonomy)

Step 1: LLM-as-a-Judge Evaluation Prepare a local LLM (ex. google/gemma-4-31b via LM Studio) to build a taxonomy for a sweep data . Open code/taxonomy_evaluator.py and check the physical parameters:

MODEL_NAME = "llama8b"       # e.g., "llama8b", "gemma7b", "llama70b"
TASK_NAME = "computer"       # e.g., "computer", "symphony"
ALPHA = 15                   # Intervention Energy Scaling Factor

This script is designed to be run from the command line. You can specify input_file and output_csv files name via command-line arguments:

python code/taxonomy_evaluator.py --input_file data/sweep_computer_llama8b_a15.jsonl --output_csv data/sweep_computer_llama8b_a15.csv

Step 2: Generate JSON for Web Open code/phenotype_json_generator.py and update the file paths to match your substrate:

# Configuration Parameters (Edit these to match your target substrate)
JSONL_FILE = "data/sweep_computer_llama8b_a15.jsonl"
CSV_FILE = "data/sweep_computer_llama8b_a15.csv"
OUTPUT_FILE = "data/heatmap_data_computer_llama_a15.json"

Then, run the generator:

python code/phenotype_json_generator.py

Step 3: Visualize Open code/heatmap.html in an editor, update const json_path to point to the newly generated JSON, and open the HTML in your browser to explore the interactive taxonomy.

// Target JSON dataset generated by phenotype_json_generator.py
const json_path = "./heatmap_data_computer_llama8b_a15.json";

🦋 Hardware and Software Environment

To ensure complete deterministic reproducibility of our continuous dynamical systems, we strictly controlled our hardware and software environments. Due to the extreme sensitivity of trajectories at the edge of chaos (as detailed in Section 4.6), specific hardware isolation was enforced.

Hardware Specifications:

• Lightweight Substrates (Llama-3.1-8B, Gemma-7B): All exploratory parameter sweeps and phase diagram generations were strictly isolated and executed on a single NVIDIA RTX Pro 4500 (Blackwell architecture) to prevent any cross-architecture floating-point divergence.
• Heavy Substrate (Llama-3.1-70B): Due to VRAM constraints imposed by the massive 70-billion parameter scale, the model was quantized to 4-bit precision (NF4) using BitsAndBytes. Inference was distributed across a heterogeneous dual-GPU setup consisting of an NVIDIA RTX Pro 4500 (Primary, CUDA:0) and an NVIDIA RTX 3090 (Ampere architecture, CUDA:1). Device-mismatch during dynamic tensor operations was prevented via real-time device alignment protocols implemented in our custom steering processor.

Software and Compilation Environment:

• Python: 3.13.14
• PyTorch: 2.10.0+cu130
• CUDA Compilation Tools: Release 13.1, V13.1.115 (Build cuda_13.1.r13.1/compiler.37061995_0) To guarantee strict deterministic reproducibility of the continuous dynamical trajectories, all pseudo-random number generators (PRNG seeds) across Python, NumPy, and PyTorch (including CUDA deterministic flags) were explicitly locked to a global seed of $42$. Furthermore, the softmax sampling temperature was strictly fixed at $0.8$ across all exploratory and scaling generations, ensuring a consistent thermodynamic baseline for the macroscopic probability field.

✍️ How to Cite

If you use our codebase, datasets, or conceptual framework in your research, please cite our arXiv preprint.

Preprint Citation (BibTeX):

@article{kayama2026semantic,
  author    = {Kayama, Yoshihiko},
  title     = {Semantic Lenia: Emergence of Homeostatic Solitons within the Semantic Space of Large Language Models},
  journal   = {arXiv preprint arXiv:2608.11657},
  year      = {2026},
  url       = {https://arxiv.org/abs/2608.11657}
}