avihayamor/tripmatch-ai-dataset

TripMatch AI Dataset

A reproducible multimodal dataset for the TripMatch AI Final Project. It contains 10,000 synthetic text trip plans with a raw idea generated for every row by the pretrained Hugging Face model google/flan-t5-small, plus 5,000 real street-view images retained as extra multimodal work. The two configurations are separate so Dataset Viewer can load each schema correctly.

Dataset statistics

Configuration	Rows	Main fields	Intended task
trip_plans	10,000	destination, duration, companions, budget, themes, daily plan	text recommendation and semantic retrieval
visual_inspiration	5,000	image, region, description, setting, weather, road type	text-to-image and image-to-image retrieval

The repository also contains three full row-aligned text embedding matrices used for the required model comparison, plus the CLIP visual matrix:

• embeddings/trip_embeddings_all_minilm_l6_v2.npy: six-layer candidate.
• embeddings/trip_embeddings_paraphrase_minilm_l3_v2.npy: winning compact candidate.
• embeddings/trip_embeddings_all_minilm_l12_v2.npy: deeper candidate.
• embeddings/visual_embeddings_clip.npy: 5,000 × 512, generated with openai/clip-vit-base-patch32.

The winner is published at avihayamor/tripmatch-ai-embedding-model.

Exploratory graphs and evaluation

All figures below are generated from the published rows and saved embeddings. They are included in this dataset repository so the dataset card shows the evidence without requiring the Space or notebook to be opened.

Visual-data coverage

The visual sample is dominated by urban/suburban, daytime and fair-weather scenes. This is an important coverage limitation rather than a claim of balanced global data.

Text-embedding clusters

K-Means model selection is performed on 645 averaged unique-destination MiniLM vectors. The selected k=11 solution has cosine silhouette 0.092, showing overlapping travel categories rather than cleanly separated natural classes.

Visual-embedding clusters

The selected CLIP solution uses k=6 with cosine silhouette 0.107. The PCA view shows broad scene structure, but substantial overlap remains.

Text recommender baselines

The production hybrid combines MiniLM cosine similarity with explicit duration, party, theme and route constraints. On 16 structured requests it outperforms the semantic-only, TF-IDF and random systems.

Usage

from datasets import load_dataset

trips = load_dataset(
    "avihayamor/tripmatch-ai-dataset",
    "trip_plans",
    split="train",
)

visuals = load_dataset(
    "avihayamor/tripmatch-ai-dataset",
    "visual_inspiration",
    split="train",
)

print(trips[0]["trip_plan"])
visuals[0]["image"].show()

Construction

Text configuration and HF synthetic generation

Licensed destination seed descriptions come from the travel portion of microsoft/RecoReact, released under CDLA-Permissive-2.0. Fifteen manually curated route archetypes were added. Structured combinations cover duration, companions, budget, pace, difficulty, season and evidence-backed themes. Then google/flan-t5-small generates a raw travel idea for every row from a constrained prompt. Each row records generation_model, generation_prompt_version, generation_decoding and generation_status.

Raw output is kept in hf_generated_idea for audit/EDA. A quality-repair layer retains the verified structured main_ideas and itinerary when the small model omits requested constraints or introduces unsupported wording—exactly the type of model failure the project instructions require students to identify. The full implementation is in scripts/generate_trip_plans_hf.py, and the executed companion notebook is notebooks/01_synthetic_data_generation.ipynb. The prompt tells the model not to invent prices, bookings, opening hours or live safety claims. The day-by-day skeleton remains deterministic so duration and exact day count are testable. These are recommendation candidates, not verified bookings or advice.

Required three-model comparison

All 10,000 final rows are embedded with:

1. sentence-transformers/all-MiniLM-L6-v2
2. sentence-transformers/paraphrase-MiniLM-L3-v2
3. sentence-transformers/all-MiniLM-L12-v2

The shared evaluation records encoding time, query latency, dimension, artifact size and top-three constraint compatibility in reports/model_comparison.csv. MiniLM-L3 is selected because it slightly improves measured compatibility over L6 and is much faster than L12 at the same 384-dimensional artifact size.

Visual configuration

This is a deterministic 5,000-row subset of Reubencf/streetview-global, pinned during construction to revision a206537534dc0e8165e0e7d36f08df14795127db. The source contains Mapillary photographs and model-generated annotations. Images here are compressed WebP thumbnails used for educational retrieval experiments; no images were generated or purchased for this project.

The subset contains 139 region labels and the following fields:

Field	Type	Description
image	Image	Compressed WebP street-view image
visual_id	string	Stable TripMatch row identifier
id	string	Original Mapillary image identifier
latitude, longitude	float64	Source coordinates
region	string	Source region label
scene_description	string	Weakly supervised scene description
setting, weather, time_of_day	string	Scene tags
road_type, infrastructure	string	Road/infrastructure tags
visual_cluster	int64	CLIP embedding cluster assignment

Quality and evaluation

• Visual rows: 5,000; unique visual_id: 5,000; null metadata cells: 0.
• CLIP text-to-image Precision@5: 0.625, versus 0.125 for the deterministic random baseline (24 labelled queries).
• CLIP image-to-image setting agreement@5: 0.731, versus 0.336 random.
• Hybrid text recommender structured-constraint score: 0.993, versus MiniLM-only 0.778, TF-IDF 0.538, and random 0.174 (16 query cases).

These small, project-specific evaluations demonstrate relative improvement; they are not a general travel benchmark.

Limitations and responsible use

• Geographic coverage is uneven and follows the source contributors.
• Most visual examples are daytime and fair-weather scenes.
• Scene descriptions and labels are weak supervision and can be wrong.
• Street imagery may contain blurred or incidentally captured people, plates, and signage.
• The trip plans are synthetic transformations of source descriptions; live prices, access, visas, weather, closures, and safety must be checked independently.
• The dataset should not be used for surveillance, identity inference, or high-stakes geographic decisions.

License and attribution

This combined repository is released under CC BY-SA 4.0. The visual configuration and its annotations inherit that license from the source. Attribute Mapillary contributors, Reuben Fernandes' StreetView Global dataset, Microsoft RecoReact, and this TripMatch dataset. The upstream RecoReact material is CDLA-Permissive-2.0; its notice is preserved in NOTICE.md.

Related resources

• App: https://huggingface.co/spaces/avihayamor/tripmatch-ai
• Model: https://huggingface.co/avihayamor/tripmatch-ai-embedding-model
• Source visual dataset: https://huggingface.co/datasets/Reubencf/streetview-global
• Source text dataset: https://huggingface.co/datasets/microsoft/RecoReact

Citation

@dataset{amor2026tripmatch,
  title        = {TripMatch AI Dataset},
  author       = {Amor, Avihay},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/avihayamor/tripmatch-ai-dataset}
}

Also cite Mapillary contributors and the two upstream Hugging Face datasets when using the corresponding configuration.