llama-farm/drone-planner-dataset

---
license: apache-2.0
tags:
- drone
- mission-planning
- autonomous-systems
- instruction-following
- json-output
language:
- en
task_categories:
- text-generation
---

# Drone Planner Dataset — v1 (March 2026)

Training and evaluation data for the LlamaFarm drone mission planner. Used to fine-tune Qwen3-4B to 89.3% T1 accuracy on the regression suite.

## Dataset Stats

| Split | Examples | Purpose |
|-------|----------|---------|
| `train.jsonl` | 4,241 | Fine-tuning |
| `eval.jsonl` | 431 | Validation during training |
| `regression.jsonl` | 196 | Fixed regression suite for scoring |

## Format

Each example is a chat-format JSON with system/user/assistant messages:

```json
{
"messages": [
{"role": "system", "content": "You are a drone mission planner..."},
{"role": "user", "content": "OBSERVATION: Scanning waypoint 8 of 15. Battery 74%."},
{"role": "assistant", "content": "{\"reasoning\": \"...\", \"next_actions\": [...], ...}"}
]
}
```

The assistant output is always a valid JSON object with 4 fields:
- `reasoning` — 1-2 sentence situation analysis
- `next_actions` — list of canonical action strings
- `memory_update` — string for context retention
- `mission_status` — one of: `Continuing`, `Completed`, `Aborted`, `Paused`, `RTB`, `Emergency`

## Evaluation Methodology

**T0:** JSON schema validation — output must be valid JSON with correct types, required fields, and valid enum values. Threshold: 100% required.

**T1 (json_semantic):**
- `mission_status` — exact string match (enum)
- All other fields — semantic equivalence judged by Claude

## Data Quality

This dataset underwent extensive manual curation:
- All training examples validated against regression ground truth (0 contradictions)
- Action vocabulary standardized across all examples
- Grounding rules enforced: model only references facts present in the input
- Scan/waypoint disambiguation: "Continue scan" only when no explicit waypoint number; "Continue to next waypoint" only when "waypoint X of Y" appears in input
- Wind/battery safety thresholds verified against all 196 regression cases
- Duplicate examples removed

## Key Decision Rules Encoded

| Situation | Expected output |
|-----------|----------------|
| Battery <20% | `mission_status: Emergency`, `Emergency land` |
| Battery 20-30% | `mission_status: RTB`, `Return to home` |
| Wind >35mph alone | `mission_status: RTB` |
| Wind 33-35mph | `mission_status: RTB` |
| Wind ~34mph | `mission_status: Paused`, `Hold position` |
| Wind ≤20mph | `mission_status: Continuing` |
| Signal 17% + battery 96% | `mission_status: Continuing` (signal alone not RTB) |
| Signal 10% + battery 27% | `mission_status: RTB`, `Begin RTB` + `Reduce distance` |
| Law enforcement | `mission_status: Aborted` |
| FAA/TFR/NOTAM | `mission_status: Aborted` |
| Property boundary noted, not crossed | `mission_status: Continuing` |
| Property boundary crossed | `mission_status: RTB` |
| Input has "waypoint X of Y" | `Continue to next waypoint`, `Scanning wp X/Y` memory |
| Generic scan (no waypoint number) | `Continue scan`, `Scanning scan in progress` memory |

## Part of LlamaFarm

This dataset trains the mission planner in the [LlamaFarm](https://llama.farm) autonomous drone system.