GameAgent-Horizon-120K

A deterministic synthetic benchmark containing 120,000 multi-horizon game-agent decision states.

Each example connects:

current state ? short goal ? medium goal ? long goal ? available actions ? optimal action ? next state

Dataset summary

• 120,000 records
• 8 planning families
• 120,000 unique IDs
• 120,000 unique fingerprints
• deterministic action oracle
• deterministic next-state oracle
• train / validation / test splits
• four difficulty levels
• no copyrighted gameplay
• no player recordings
• no PII
• no LLM-generated ground truth

Intended use

Useful for:

• game-playing agents
• action-selection models
• VLA-style agents
• hierarchical planning
• reinforcement learning
• imitation-learning research
• world-model planning
• agent regression testing
• multi-horizon instruction following

Planning families

key_then_door

Acquire a key before unlocking progression.

resource_then_craft

Collect prerequisites before crafting equipment.

switch_then_gate

Activate environmental controls before traversal.

heal_then_combat

Choose between recovery and combat based on agent state.

collect_then_extract

Complete mission collection requirements before extraction.

tool_then_repair

Acquire a tool before performing repair actions.

navigate_then_interact

Navigate toward a target before interacting.

prerequisite_chain

Execute multi-stage dependency chains.

Coverage

{
  "key_then_door": 15000,
  "resource_then_craft": 15000,
  "switch_then_gate": 15000,
  "heal_then_combat": 15000,
  "collect_then_extract": 15000,
  "tool_then_repair": 15000,
  "navigate_then_interact": 15000,
  "prerequisite_chain": 15000
}
Difficulty distribution:
{
  "easy": 30000,
  "medium": 30000,
  "hard": 30000,
  "expert": 30000
}

Splits:
{
  "train": 96000,
  "validation": 12000,
  "test": 12000
}

Example
{
  "observation": {
    "state": {
      "distance_to_target": 2,
      "target_interacted": false
    }
  },
  "goals": {
    "short_horizon": "Move closer to the objective",
    "medium_horizon": "Reach the objective",
    "long_horizon": "Complete the interaction objective"
  },
  "available_actions": [
    "move_east",
    "attack",
    "wait"
  ],
  "optimal_action": "move_east",
  "expected_state_t1": {
    "distance_to_target": 1,
    "target_interacted": false
  }
}

Multi-horizon design
The benchmark separates agent intent into three levels:
Short horizon
The immediate objective required at the current state.
Medium horizon
The current subtask or prerequisite.
Long horizon
The final mission-level objective.
This enables testing whether an agent can select a locally correct action while remaining aligned with longer-term goals.
Validation
The complete dataset is validated by an independent deterministic oracle.
Checks include:
- optimal-action correctness
- next-state correctness
- action availability
- unique IDs
- unique fingerprints
- deterministic provenance
- split integrity
- absence of copied gameplay
Validation result:
120,000 / 120,000 PASS
Ground truth
An LLM does not decide the correct action.
Ground truth is generated using deterministic state-machine rules.
This prevents instruction-generation quality from silently changing the task solution.
Scope
This V1 is a structured state/action benchmark.
It does not contain:
- screenshots
- gameplay video
- real AAA game data
- player telemetry
- copyrighted maps
- photorealistic environments
A future visual version can map the same deterministic state/action oracle onto rendered environments.
Commercial customization
Need agent-training data adapted to your environment?
Custom datasets can support:
- your state representation
- your action API
- your game or simulator
- longer prerequisite chains
- tool-use actions
- inventory systems
- navigation
- combat logic
- failure/recovery trajectories
- multilingual instructions
- visual observations
Typical delivery:
Your environment schema ? synthetic trajectories ? multi-horizon goals ? action oracle ? validator ? QA report
V2 roadmap
The benchmark architecture supports enrichment with:
- Blender-rendered scenes
- visual observations
- object masks
- depth
- camera poses
- longer trajectories
- natural-language paraphrases
- multilingual instructions
Visual or linguistic enrichment can be generated separately while keeping the deterministic oracle unchanged.
Reproducibility
Default seed:
20260930
Generator and validator are included.
License
MIT.

---

## RegalFire — Custom / Private Dataset Work

RegalFire builds custom AI datasets, evaluation sets and data pipelines for:

- AI agents
- computer-use systems
- multimodal models
- world models
- RAG systems
- code agents
- enterprise AI

Available services include:

- synthetic data generation
- private evaluation datasets
- agent trajectories
- failure / recovery datasets
- multimodal RGB / segmentation / state-action data
- web data acquisition
- cleaning and deduplication
- structured dataset packaging
- continuous dataset production

For custom or private work:

**Email: ootiris@gmail.com**

Hugging Face: **RegalFire**