---
license: mit
language:
- en
tags:
- artificial-life
- reinforcement-learning
- world-model
- dl2l
task_categories:
- other
---

# DL2L Experiments Dataset

Simulation trajectory data from the [DL2L](https://github.com/felipedreis/dl2l)
distributed artificial life simulator, used to train JEPA world models.
See [`felipedreis/dl2l-jepa`](https://huggingface.co/felipedreis/dl2l-jepa) for the trained models.

## Dataset structure

Data is organized by experiment prefix. Each prefix contains parquet files for
model training and a `stats.json` with dataset metadata.

```
p9/
  train.parquet       # single-encoder training set (trials 1–8)
  val.parquet         # single-encoder validation set (trials 9–10)
  train_dual.parquet  # dual-encoder training set (adds h_t homeostatic columns)
  val_dual.parquet    # dual-encoder validation set
  stats.json          # dims, feature order, normalisation stats, split sizes
```

## p9 experiment

**Simulation:** 10 creatures × 10 trials, DL2L `basic.conf` world.  
**Split:** trials 1–8 → train, trials 9–10 → val (trial-based to prevent cross-trial contamination).  
**Sizes:** 359,782 train / 89,731 val samples.

### Sample format

Each row is a `(s_t, a_t, emotion_target)` tuple:

| Column group | Columns | Description |
|---|---|---|
| Perception `s_t` | `distance`, `angle`, `direction` | Target object spatial features |
| Object type | `type_GRAY_APPLE`, `type_GREEN_APPLE`, `type_RED_APPLE`, `type_ROTTEN_APPLE`, `type_CACTUS`, `type_ALOE` | One-hot object type |
| Action `a_t` | `a_APPROACH`, `a_AVOID`, `a_EAT`, `a_ESCAPE`, `a_PLAY`, `a_SLEEP`, `a_TOUCH`, `a_TURN`, `a_WANDER` | One-hot chosen action |
| Target emotion | `final_hunger`, `final_sleep`, `final_apathy`, `final_stress`, `final_pain`, `final_tedium`, `final_fear`, `final_curiosity`, `final_fertility` | Absolute arousal after next regulation |

Dual-encoder parquet files additionally include:

| Column | Description |
|---|---|
| `ht_hunger`, `ht_sleep`, `ht_pain`, `ht_tedium` | Homeostatic state at action time (`h_t`) |

### stats.json

```json
{
  "input_dim": 9,
  "action_dim": 9,
  "emotion_dim": 9,
  "latent_dim": 64,
  "internal_state_dim": 4,
  "internal_latent_dim": 16,
  "live_emotion_dims": [0, 1, 4, 5],
  "perception_feature_order": ["distance", "angle", "direction", ...],
  "action_index_order": ["APPROACH", "AVOID", "EAT", "ESCAPE", "PLAY", "SLEEP", "TOUCH", "TURN", "WANDER"],
  "emotion_index_order": ["hunger", "sleep", "apathy", "stress", "pain", "tedium", "fear", "curiosity", "fertility"],
  "feature_means": [...],
  "feature_stds": [...],
  "n_train": 359782,
  "n_val": 89731
}
```

## Data extraction

Raw data is extracted from the PostgreSQL database with:

```bash
python3 scripts/pg_extract.py --out /path/to/output --container <db-container>
```

This covers trajectories, sleep episodes, engrams, arousal history, behavioural
efficiency, perception coverage, traveled distances, consolidation batch stats,
and more. See `scripts/pg_extract.py` for the full list.

The ML training dataset is assembled from the CSV output with:

```bash
cd ml
python3 -m scripts.prepare_dataset --wd /path/to/output --out data_p9 --dual
```

## Citation

```
DL2L — Distributed Live to Learn, Learn to Live
Felipe Duarte dos Reis, CEFET-MG, 2017–2026
https://github.com/felipedreis/dl2l
```