Datasets:

tostido
/

key-data

pretty_name: KEY Neuroevolution Dataset
license: mit
tags:
  - neuroevolution
  - lora
  - genetic-algorithms
  - provenance
  - world-model
language:
  - en
configs:
  - config_name: comm_events
    data_files:
      - split: train
        path: data/comm_events/train.jsonl
  - config_name: crossovers
    data_files:
      - split: train
        path: data/crossovers/train.jsonl
  - config_name: selection
    data_files:
      - split: train
        path: data/selection/train.jsonl
  - config_name: mutations
    data_files:
      - split: train
        path: data/mutations/train.jsonl
  - config_name: fitness
    data_files:
      - split: train
        path: data/fitness/train.jsonl
  - config_name: performance
    data_files:
      - split: train
        path: data/performance/train.jsonl
  - config_name: errors
    data_files:
      - split: train
        path: data/errors/train.jsonl
  - config_name: evolution_events
    data_files:
      - split: train
        path: data/evolution_events/train.jsonl

🔑 KEY: Neuroevolution Dataset

40,000+ logged events from real evolutionary runs — every mutation, crossover, selection, and fitness evaluation.

KEY evolves LoRA adapters on frozen base models (MiniLM-L6, DreamerV3) using NEAT-style neuroevolution. This dataset captures the complete evolutionary history.

🎮 Links


🌌 Live Demo	Watch evolution in action
🧠 Champion Model	The evolved DreamerV3 model

Loading the Dataset

from datasets import load_dataset

# Available configs:
ds = load_dataset("tostido/key-data", "comm_events")      # 16,968 rows - pod communication
ds = load_dataset("tostido/key-data", "crossovers")       # 8,878 rows - breeding events
ds = load_dataset("tostido/key-data", "selection")        # 4,266 rows - tournament selection
ds = load_dataset("tostido/key-data", "mutations")        # 3,848 rows - mutation events
ds = load_dataset("tostido/key-data", "fitness")          # 2,121 rows - fitness evaluations
ds = load_dataset("tostido/key-data", "performance")      # 2,121 rows - runtime telemetry
ds = load_dataset("tostido/key-data", "errors")           # 2,070 rows - errors/warnings
ds = load_dataset("tostido/key-data", "evolution_events") # event bus stream

Example: Evolving Semantic Similarity

Task: Adapt MiniLM embeddings to preserve semantic relationships

Test Pair: "The cat sat on the mat" ↔ "A feline rested on the rug"

Generation	Cosine Similarity	Fitness
0	0.42 (random)	0.35
50	0.76	0.64
100	0.89	0.82

The evolved adapter learned to preserve semantic similarity while improving output quality.

What Gets Evolved

KEY freezes the base model and evolves only the adapter:

┌──────────────────────────────────────┐
│         Evolvable Brain              │
│  ┌────────────────────────────────┐  │
│  │   Base Model (FROZEN)          │  │  ← MiniLM (22M) or DreamerV3 (200M)
│  └─────────────┬──────────────────┘  │
│                ▼                     │
│  ┌────────────────────────────────┐  │
│  │   LoRA Adapter    (~12K)       │  │  ← EVOLVED
│  │   Projection Head (~99K)       │  │  ← EVOLVED
│  └────────────────────────────────┘  │
└──────────────────────────────────────┘

Total evolved parameters: ~111K (vs 22M-200M frozen)

Fitness Functions

What evolution optimized for (from fitness.jsonl):

AdapterFitness (Interface Quality)

Preservation (40%): Does adapter maintain semantic structure?
Signal Quality (30%): Is output well-conditioned? (not collapsed/exploded)
Consistency (30%): Similar inputs → similar outputs?

EmbeddingKleeneFitness (Semantic Convergence)

Coherence: Similar pairs should have high cosine similarity
Separation: Dissimilar pairs should be far apart
Convergence: Embedding variance stays bounded

DreamerFitness (World Model Quality)

Prediction: How well does imagination match reality?
Stability: Do trajectories stay bounded?
Reward: Can the model anticipate outcomes?

Schema Reference

`mutations.jsonl`

{
  "timestamp": 1737403521.234,
  "event": "mutation",
  "generation": 42,
  "parent_id": "node_abc123",
  "child_id": "node_def456",
  "parent_fitness": 0.72,
  "mutation_rate": 0.1,
  "mutated_traits": ["exploration", "caution"],
  "deltas": {"exploration": 0.05, "caution": -0.02}
}

`crossovers.jsonl`

{
  "event": "crossover",
  "generation": 42,
  "parent1_id": "node_abc",
  "parent2_id": "node_xyz",
  "child_id": "node_new",
  "parent1_fitness": 0.72,
  "parent2_fitness": 0.68,
  "contribution_p1": 0.55
}

`fitness.jsonl`

{
  "event": "fitness_evaluation",
  "generation": 42,
  "node_id": "node_abc123",
  "fitness_function": "AdapterFitness",
  "raw_fitness": 0.823,
  "components": {
    "preservation": 0.85,
    "signal": 0.79,
    "consistency": 0.84
  },
  "eval_time_ms": 45.2
}

`selection.jsonl`

{
  "event": "selection",
  "generation": 42,
  "method": "tournament",
  "survivors": ["node_a", "node_b", "node_c"],
  "eliminated": ["node_d", "node_e"],
  "elites_preserved": 2
}

Why Evolve Instead of Gradient Descent?

Neuroevolution works when:

✅ Your objective isn't differentiable (human preference, discrete outputs)
✅ You want population diversity (speciation prevents local optima)
✅ You're optimizing for interface quality, not task loss
✅ You need full auditability (every mutation logged with provenance)

FAQ

Q: What's a "quine brain"?

A brain that can serialize its weights → mutate → deserialize. This enables genetic algorithms to evolve neural networks. Think "self-modifying adapter."

Q: Why not just use backprop?

Backprop requires differentiable objectives. Evolution works with any fitness function: human ratings, game scores, discrete metrics.

Q: Is this real data?

Yes. This dataset contains 40K+ events from actual evolutionary runs.

🔐 Get Full Source Access

Tier	Price	What You Get
🔑 Source Access	$100 one-time	Full codebase, private repo invite
🤝 Hands-On	$50/hour	I coach you through wiring your own model
🛠️ Done-For-You	$500 flat	I wire up your custom model for you
🎤 Speaking	$2,000	Talk at your company on gradient-free optimization