Synthetic Packing Lists Dataset (Syn-Item-Lists)

A synthetic, expert-quality dataset of travel packing lists designed specifically for fine-tuning travel logistics assistants, recommendation systems, and structured-output LLMs.

Manually curating comprehensive packing lists that adapt to highly specific travel conditions (like climate, infrastructure, and active risks) is a massive bottleneck. To solve this, this dataset was generated using knowledge distillation via Google Gemini, grounded tightly on 10 hand-curated, gold-standard seed examples.

Unlike naive synthetic data, this dataset utilizes a hierarchical sampling framework to ensure that every scenario is logistically realistic and structurally sound.

🔗 External Resources

• Hugging Face Repository: rinad12/Syn-Item-Lists

📐 Dataset Structure & Schema

The dataset follows a standard 80/10/10 split distributed across three files:

• train.jsonl (80%)

• val.jsonl (10%)

• test.jsonl (10%)

Each line is a self-contained JSON record consisting of an input travel context and a flat output list of recommended items.

Input Schema (The Context)

Field	Type	Description
intent	String	Trip purpose (e.g., "Hiking / Trekking", "Wildlife safari", "Business trip")
duration	Integer	Trip length in days
infrastructure	String	Availability of utilities/services at the destination (e.g., "Basic — rural / remote village")
climate	String	Climate type at the destination (e.g., "Subtropical")
risks	List[String]	1–3 active risk factors (e.g., "High UV exposure", "Vector-borne disease")
max_items	Integer	Upper bound constraint on the list length passed to the model

Output Schema (The Packing List)

Field	Type	Description
item	String	Highly specific item name (e.g., "Merino Wool Socks", instead of just "Socks")
quantity	Integer	Number of units recommended to pack
reason	String	Concise justification of why the item is needed (max 10 words)

JSON Example

{
  "input": {
    "intent": "Wildlife safari",
    "duration": 10,
    "infrastructure": "Basic — rural / remote village",
    "climate": "Subtropical (hot humid summer)",
    "risks": ["Vector-borne disease (mosquitoes / ticks)", "High UV exposure"],
    "max_items": 13
  },
  "output": [
    {
      "item": "Long-Sleeved Lightweight Shirt",
      "quantity": 5,
      "reason": "Barrier against mosquitoes at dawn/dusk drives."
    },
    {
      "item": "DEET Insect Repellent",
      "quantity": 2,
      "reason": "Essential protection against vector-borne disease."
    }
  ]
}

⚙️ How the Data Was Generated

To maintain exceptionally high data quality and avoid "hallucinated" or nonsensical travel scenarios, the generation pipeline utilizes several strict guardrails:

1. Hierarchical Dependency Sampling: Scenarios are constructed top-down: Intent → Infrastructure → Climate → Risks. This ensures you never see impossible combinations, like a corporate business trip requiring a tent in an Arctic tundra.

2. Intent-Aware Durations: Trip lengths are drawn from logical, intent-specific ranges (e.g., Business trips: 1–5 days; Digital Nomad stays: 14–90 days).

3. Dynamic List Caps: The max_items field is mathematically constrained using a base duration tier, modified by the intent type (e.g., Camping adds items, Business subtracts items), and padded with a risk floor.

4. Generalization Testing (Seen/Unseen Splits): The validation and test sets are carefully split—50% feature context combinations seen during training, and 50% feature completely novel combinations. This allows you to strictly measure how well your fine-tuned model generalizes to unseen travel environments.

5. Strict Schema Enforcement: Every sample was generated using Gemini's native response_schema leveraging Pydantic constraints, meaning zero malformed JSON outputs.

💡 Potential Use Cases

• LLM Fine-Tuning: Ideal for instruction-tuning smaller open-source models (like LLaMA, Mistral, or Qwen) to handle highly constrained, structured JSON generation tasks.

• Travel Recommendation Systems: Training classification or embedding models to suggest hyper-specific gear based on user-input metadata.

• RAG Evaluation: Testing retrieval-augmented generation pipelines on how well they adjust recommendations to dynamic risk and climate factors.

📜 License

This dataset is open-source and licensed under the Apache 2.0 License. Feel free to use it for both commercial and research purposes!