EpicVic2193/restaurant-inventory-forecasting

🍽️ Restaurant Inventory Forecasting Dataset

Overview

High-quality synthetic training data for post-training a small LLM (Qwen2.5-1.5B) to become an expert reasoning engine for restaurant inventory forecasting.

Dataset Structure

Split	Examples	Purpose
sft_train	~23,750	SFT training (instruction-response pairs with CoT reasoning)
sft_eval	~1,250	SFT evaluation
grpo_train	~4,500	GRPO RL training (prompts + ground truth solutions)
grpo_eval	~500	GRPO evaluation

Data Format

SFT Split (ChatML messages format)

{
  "messages": [
    {"role": "system", "content": "You are a precise restaurant inventory forecasting AI..."},
    {"role": "user", "content": "Item: Pistachio Kunafa... Historical sales: [40, 45, 50]... Tomorrow: Friday, rainy..."},
    {"role": "assistant", "content": "<think>\nStep 1: Historical baseline...\nStep 2: Day-of-week pattern...\n</think>\n<answer>\n{\"item\": \"Pistachio Kunafa\", \"predicted_demand\": 52, ...}\n</answer>"}
  ]
}

GRPO Split (prompt + solution format)

{
  "prompt": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}],
  "solution": "{\"item\": \"Pistachio Kunafa\", \"predicted_demand\": 52, \"recommended_order\": 60, \"unit\": \"portions\"}"
}

Domain Coverage

• 7 restaurant types: Middle Eastern Café, Italian Trattoria, Japanese Ramen Shop, American Burger Joint, Mexican Taqueria, Indian Curry House, French Bakery
• 40+ menu items across categories: main, appetizer, dessert, hot_beverage, cold_beverage, side, pastry
• 8 weather conditions: sunny, cloudy, rainy, stormy, snowy, hot, cold, mild
• 12 special events: festivals, holidays, sports games, etc.
• 9 location types: downtown, university, mall, tourist area, etc.

Scenario Types

Scenario	% of Data	Description
Standard	40%	Normal day-to-day forecasting
Weather shift	15%	Sudden weather change affecting demand
Multi-item	10%	Predict 3-5 items simultaneously
Demand spike	10%	Event-driven demand surges
Stockout risk	10%	Upward trending demand with stockout alerts
Slow day	10%	Low-traffic day combinations
Heatwave	5%	Extended hot weather patterns

Training Recipe

1. SFT Stage: Fine-tune on sft_train split with completion-only loss
2. GRPO Stage: Reinforce with grpo_train using MAPE-based reward function
3. Evaluation: Test on held-out splits with MAPE metric

Citation

Generated using domain-informed simulation engine based on real restaurant industry patterns. Methodology inspired by:

• DeepSeekMath (arXiv:2402.03300) — SFT→GRPO pipeline
• Chronos (arXiv:2403.07815) — KernelSynth synthetic data approach
• LLMTime (arXiv:2310.07820) — Time-series text encoding
• Context is Key (arXiv:2410.18959) — Text-augmented forecasting