Turhan123/astra-meal-parser

Astra Meal Parser: Structured Meal Data Extraction

Turhan123/astra-meal-parser is a specialized 3.1 billion parameter model, fine-tuned from Qwen2.5-3B-Instruct by Turhan Göksu. Its core function is to accurately extract food items and their corresponding amounts from natural language meal descriptions, supporting both Turkish and English, as well as mixed-language inputs.

Key Capabilities

• Structured Output: Generates a strict JSON object containing name and amount for each food item, without any additional prose or macros.
• Bilingual & Code-Switched Input: Handles inputs like "200g grilled chicken ve 1 kase pirinç" seamlessly.
• High Parsing Accuracy: Achieves 100% precision, recall, and F1 score for item parsing on its evaluation set, with zero parse failures.
• Optimized for Nutrition Pipelines: Designed to feed into deterministic nutrition calculators, significantly reducing calorie estimation errors (down to 3.1% MAPE for the full pipeline) by separating natural language parsing from numerical calculation.

Good For

• Meal Logging Applications: Ideal for calorie tracking or dietary apps where users input meals in free text.
• Front-end for Nutrition Calculators: Provides a robust parsing layer for systems that combine natural language input with a static nutrition database.

Limitations

• Parsing Only: This model does not estimate calories or macros directly; it only extracts items and amounts. Accuracy of nutrition data depends on the external nutrition table and calculator.
• Portion Ambiguity: Vague amounts (e.g., "1 bowl of rice") are resolved with default serving sizes, which can be a source of residual error.
• Strict JSON Output: While highly robust, consuming applications should still implement safeguards for occasional malformed responses.