FerasMad/arabic-restaurant-ambience

Arabic Restaurant Ambience Annotations

A dataset for detecting ambience / physical-place complaints in Arabic restaurant reviews — the kind of complaint that's about the room itself (AC, seating, noise, smell, decor, parking) rather than the food, staff, delivery, or price.

Built as part of the Arabic Restaurant Complaints Classifier project. Released because there's almost no public Arabic ambience-labeled data, and we wanted to fix that.

What's in here

File	Rows	What it is	License
synthetic_contrastive_pairs.csv	250	125 hand-written contrastive pairs across 12 ambience subtypes. Each pair shares a "risky" word (بارد, حار, قديم, ريحة, زيت) but one row is ambience and the other is the matching non-ambience hard-negative.	MIT (ours)
adversarial_test_set.csv	183	Hand-written adversarial test cases across 13 attack types: clean, boundary, negation, sarcasm, mixed dialect, multi-aspect, very_short, long, out_of_domain, typo, emoji, adversarial.	MIT (ours)
boundary_examples.csv	90	Hand-written boundary cases that map to the canonical labeling rules in docs/BOUNDARIES.md.	MIT (ours)
ambience_keywords.json	–	Saudi/Gulf ambience keyword vocabulary across 12 subtypes + risky-ambiguous tokens + 7 hard-negative keyword sets (food, service, delivery, price, order, hygiene, wait).	MIT (ours)
annotations_qaym_*.csv	3,882 total	Filter outputs on hadyelsahar/ar_res_reviews. Annotations only — no source text. Re-join with the source dataset by id.	Annotations: MIT (ours). Source text: see hadyelsahar's dataset card.
annotations_hard_*.csv	41,752 total	Filter outputs on HARD (hotel reviews — useful for transfer to restaurant ambience). Annotations only — no source text. Re-join by id.	Annotations: MIT (ours). Source text: see HARD's dataset card.
complaints_labeled.csv	95,391	The full production training dataset. Text + 8-class labels + source + priority. The actual data the production model trained on. Sources: Saudi delivery-app reviews (production, ~87K), team-generated synthetic templates (synthetic, ~7K), Play Store reviews (play_store, ~1.1K), one external source (res1, ~570). Released for research / academic / non-commercial use only — the labels are 100% the AI Club NLP team's manual work; the underlying review text comes from public Saudi delivery-app reviews. If you need commercial use of the source text, contact the original platforms directly.	Labels + curation: MIT. Source text: research-use only.
production_labels_only.csv	95,391	The 8-class production labels — the actual training labels for the deployed model. Columns: row_id, category, source, priority, text_sha256 (16-char prefix for de-dup), text_len_chars, text_len_words. No source text. Source attribution: collected from public Saudi delivery-app reviews + synthetic templates + scraped Play Store reviews — text redistribution rights uncertain, so labels-only is the safe path.	MIT (the labels are 100% the team's manual work). Source text: not redistributed.
docs/BOUNDARIES.md	–	The exact labeling rules. The canonical "what is and isn't ambience" reference.	MIT
docs/DATA_SOURCES.md	–	Tier 1/2/3 source ranking + ethical collection guidance.	MIT
docs/DATA_SCHEMA.md	–	19-column schema for ambience-experiment data files.	MIT

The 12 ambience subtypes

temperature_ac        seating_comfort       space_crowding
noise_music           lighting              smell
decor_furniture       parking               bathroom_facilities
outdoor_view          privacy               general_place_vibe

Each ambience-positive row in synthetic_contrastive_pairs.csv and boundary_examples.csv has an ambience_subtype column tagging which of the 12 it belongs to.

Why this exists

The 8-class production model at the parent project intentionally dropped the ambience class in v3 → v4 because labels were ~99% noise. v5 is the experimental retry. The boundary rules are explicit, the keyword vocab is restaurant-grounded, and the synthetic pairs deliberately train the model on what makes ambience different from food temperature, food smell, dirty surfaces, and staff behavior.

If you're building an Arabic aspect-classification model, the boundary cases are where you'll fail. We documented every one we hit.

Recommended use

For training a 9-class Arabic restaurant complaint classifier that includes ambience:

• Use synthetic_contrastive_pairs.csv as supervised training data (training-only — never in val/test, by source rule)
• Use annotations_qaym_*.csv as a filtering layer on hadyelsahar/ar_res_reviews to identify weak-labeled real ambience examples
• Use boundary_examples.csv as a known-good test set for boundary cases
• Use adversarial_test_set.csv as the hard-test gate before shipping

For continued pre-training (MLM) on Arabic restaurant text:

• Use annotations_qaym_*.csv and annotations_hard_*.csv id columns to fetch source text from the original datasets, then run MLM on that text

For evaluation only:

• Use adversarial_test_set.csv as a held-out test set for any Arabic restaurant ambience model. Don't include it in your training data.

Loading

from datasets import load_dataset

# Load the synthetic contrastive pairs
ds = load_dataset("FerasMad/arabic-restaurant-ambience", data_files="synthetic_contrastive_pairs.csv")

Or just pd.read_csv — these are plain CSVs with UTF-8 BOM (Excel-friendly).

Reproducing the harvested annotations

See docs/REPRODUCE.md for step-by-step instructions to recover the full text + annotations by joining annotations_*.csv against the upstream HF datasets.

Boundary rules (the most important section)

Three rules that determine ambience labeling — full table in docs/BOUNDARIES.md:

Rule	Ambience	Not ambience
Temperature	المكان حار والمكيف ما يبرد (room temp / AC)	الأكل بارد (food temp)
Smell	ريحة الزيت في المطعم (environment)	ريحة الأكل غريبة (food)
Furniture	الطاولة تهتز (broken/uncomfortable)	الطاولة وصخة (dirty → cleanliness)
Bathroom	السيفون خربان (broken facility)	الحمام وصخ (dirty → cleanliness)
Staff vs ambience	n/a	الموظف بارد (staff behavior, not room temp)

If your model can label all 12 boundary examples in boundary_examples.csv correctly, it understands ambience.

Performance benchmarks

The associated model (CAMeLBERT-mix fine-tuned, available at FerasMad/arabic-complaints-classifier) achieves on adversarial_test_set.csv:

• Ambience F1: 89.22% (with threshold tuning + abstain wrapper)
• Ambience precision: 86.67%
• Ambience recall: 91.92%
• Per-attack-type pass rates: clean 89%, boundary 80%, sarcasm 100%, long 87%, multi_aspect 40%, OOD 50%

Multi-aspect remains the hardest case — single-label classification can't decompose "الاكل بارد والمكيف خربان" cleanly. Multi-label retraining is the next step.

Citation

If you use this dataset, please cite:

@misc{madkhali2026ambience,
  title  = {Arabic Restaurant Ambience Annotations},
  author = {Madkhali, Feras and the AI Club NLP Team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/FerasMad/arabic-restaurant-ambience},
  note   = {Annotations on hadyelsahar/ar_res_reviews and HARD; synthetic contrastive pairs and adversarial test set are original work}
}

Also cite the upstream datasets if you use the source text:

• hadyelsahar/ar_res_reviews — Arabic restaurant reviews from qaym.com
• HARD — Hotel Arabic Reviews Dataset

License

• Annotations + synthetic + adversarial + keyword vocab + docs: MIT (your work)
• Source text (when joined back from upstream datasets): per the upstream dataset's license — read each card before redistribution

Contributing

Issues, label corrections, and contributions are welcome at the parent repo: https://github.com/FerasMad/NLP-complaints-system

If you label more rows in any of the candidate sets, please share back so the community gets cleaner training data over time.

---

Built by: the NLP team at AI Club, led by Feras Madkhali. Built with: ChatGPT, Claude (Anthropic), Codex (OpenAI) — used for boilerplate and analysis. Schema design, boundary rules, keyword curation, and labeling decisions are the team's work.