OpenDB_Compliments_Merged/README.md

---
language:
- en
- fr
license: mit
task_categories:
- text-classification
- tabular-classification
tags:
- grocery
- deduplication
- retail
- food
pretty_name: Bilingual Grocery Product Deduplication
---

# Bilingual Grocery Product Deduplication Dataset

## Dataset Summary
This dataset contains a fully deduplicated, canonical list of bilingual (English and French) grocery products. The original raw data consisted of noisy grocery inventory records with varying formats, typos, languages, and branding.

This repository provides the final cleaned dataset, the mapping of raw barcodes to canonical IDs, and a side-by-side comparison file to review the deduplication results. The deduplication was performed using a high-precision "Name-First" architecture combining token-overlap heuristics, multilingual embeddings, and Large Language Model (LLM) reasoning.

## File Structure

| File | Description |
|---|---|
| `canonical_products.parquet` | The clean, deduplicated database of the 1,109 unique canonical products, complete with averaged nutritional data. |
| `product_mapping.parquet` | The lookup table mapping every original 13-digit barcode (`code`) to its new `canonical_product_id`. |
| `deduplicated_comparison.csv` | A flat, human-readable CSV that joins the original data side-by-side with the new canonical groupings. Excellent for human review of the "before and after" state. |
| `Deduplication_Pipeline_Documentation.docx` | Comprehensive documentation outlining the v3 pipeline architecture, the modifier vetoes, the LLM usage, and final metrics. |
| `preprocess.py` | Phase 1 of the pipeline: Strict text normalization, cleaning, and exact-matching. Drops unnamed rows. |
| `candidate_gen.py` | Phase 2 of the pipeline: Generates fuzzy-match candidate pairs using a 0.65 Jaccard overlap threshold, bilingual embeddings, and applies hardcoded Modifier/Ingredient Vetoes. |
| `llm_grouper.py` | Phase 3 of the pipeline: Feeds the surviving ambiguous clusters to the LLM for atomic, semantic grouping to prevent chain-merging. |
| `llm_utils.py` | A custom LLM load balancer that routes requests between Groq, OpenRouter, NVIDIA, and Ollama, handling rate limits and JSON parsing. |
| `generate_outputs.py` | Phase 4 of the pipeline: Reads the final canonical JSON and builds the resulting Parquet and CSV dataset files. |

## Methodology (The v3 "Name-First" Architecture)

The dataset was constructed using a highly defensive deduplication pipeline designed to prioritize precision over recall (i.e., strictly avoiding the merging of different flavours, heat levels, or base ingredients).

1. **Phase 1: Strict Normalization & Exact Match**
   - Stripped brand prefixes (e.g., "Compliments", "Sensations").
   - Fixed French encoding artifacts and normalized casing/plurals.
   - Grouped all exact textual matches automatically. Unnamed barcode rows were intentionally dropped from the deduplication pool to prevent nutrition-based pollution.

2. **Phase 2: High-Confidence Candidate Generation & Vetoes**
   - Pairs were generated using a strict **Jaccard Token Overlap** threshold (≥ 0.65). 
   - Multilingual embeddings (`paraphrase-multilingual-MiniLM-L12-v2`) were selectively applied *only* to cross-language (English ↔ French) pairs.
   - **Modifier Veto:** Before any candidate was approved, it passed through a rigorous check against hardcoded lists of flavours, heat levels, fat percentages, and core ingredient nouns. E.g., If a candidate pair consisted of *Cherry Jelly Powder* and *Lime Jelly Powder*, the flavour mismatch triggered an auto-veto.

3. **Phase 3: Group-Based LLM Judging**
   - Surviving ambiguous candidates were clustered into connected components.
   - These small clusters were passed to an LLM (Llama 3.3 70B / Llama 3.1 8B). The LLM evaluated entire groups atomically to explicitly prevent transitive chain-merging (where A=B and B=C incorrectly causes A=C).

## Results
- **Original Dataset:** 2,078 rows (1,322 named products, 756 unnamed)
- **Final Canonical Products:** 1,109
- **Safe Deduplication Rate:** 46.6% (on named products)

The pipeline successfully resolved complex typos (e.g., *Two Bite Brownines* -> *TWO-BITE Brownies*) and bilingual equivalents (e.g., *Fromage Cottage* -> *Cottage Cheese*), while successfully isolating specific product varieties (e.g., separating *Lemon Lime Sparkling Water* from *Orange Sparkling Water*).