Retail-TW-Qwen3-Embedding-0.6B

Retail-TW-Qwen3-Embedding-0.6B is a text embedding model fine-tuned from Qwen/Qwen3-Embedding-0.6B for product retrieval in Taiwan retail and e-commerce. It encodes shopper queries and product listings (title, description, and category) into a shared vector space for semantic search.

Model Details

Key Features

• 🛒 Retail-Domain Adaptation: Fine-tuned on Taiwan e-commerce product data and shopper queries, covering brand searches, gift-oriented queries, vague product recollections, and typographical errors.
• 🎯 Hard-Negative Training: Trained against mined hard negatives, i.e., look-alike products that share a brand or category but differ in model, color, or pack size, to improve fine-grained product discrimination.
• 🇹🇼 Traditional Chinese Focus: Designed for zh-TW queries, with support for English brand and product names.
• 📐 Flexible Embedding Dimensions: Embeddings can be truncated to reduce storage and latency, with limited loss in retrieval quality down to 384 dimensions.

Training Data

The training data is proprietary Taiwan retail data drawn from PIC's e-commerce product catalog (title, description, and category). Training queries were synthetically generated, and hard negatives were mined from the same catalog.

Training Procedure

Evaluation

Evaluation Setup

The fine-tuned model is compared against its base model, Qwen/Qwen3-Embedding-0.6B, on an internal retrieval benchmark of 500 retail products and 1,000 LLM-generated shopper queries. The same queries are used for both models. All metrics are higher-is-better. No significance test was performed; the markers indicate the direction of change only:

🟢 improvement · 🔴 degradation · ⚪ no change

Note: The benchmark is small and uses synthetic queries over a limited catalog, so absolute scores are high and differences between models are compressed. Results indicate relative improvement over the base model and may not transfer to other catalogs.

Retrieval Quality

Metrics: Recall@1, Recall@5, MRR@10, and NDCG@10.

BaselineFine-tuned modelBaselineTunedΔ PP Recall@189.3%95.4%+6.1% Recall@597.7%99.9%+2.2% MRR@1092.9%97.4%+4.5% NDCG@1094.4%98.0%+3.6%

Embedding Dimensions

Retrieval quality after truncating the embedding to a smaller dimension, evaluated on the same benchmark.

Recall@1

BaselineFine-tuned modelBaselineTunedΔ PP 1024 dimensions89.3%95.4%+6.1% 768 dimensions89.5%95.4%+5.9% 512 dimensions88.2%95.4%+7.2% 384 dimensions87.2%95.1%+7.9% 128 dimensions81.3%92.7%+11.4%

MRR@10

Baseline values are not available for this metric.

Fine-tuned modelTuned 1024 dimensions97.4% 768 dimensions97.4% 512 dimensions97.2% 384 dimensions97.0% 128 dimensions95.2%

NDCG@10

Baseline values are not available for this metric.

Fine-tuned modelTuned 1024 dimensions98.0% 768 dimensions98.0% 512 dimensions97.9% 384 dimensions97.7% 128 dimensions96.1%

Usage

Input Format

The model was trained with a task instruction on the query side and a Document: prefix on the product side. The same format should be used at inference.

Sentence Transformers

pip install -U sentence-transformers transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("UNI-PIC/Retail-TW-Qwen3-Embedding-0.6B")

instruction = (
    "Instruct: 尋找符合消費者意圖的商品。商品必須是查詢的目標物品，而非僅含相關成分或品牌。"
    "請考慮：品牌搜尋、送禮意圖、模糊記憶、錯拼搜尋等台灣消費者真實搜尋行為。\n"
    "Query: "
)

queries = [instruction + q for q in ["洗髮精", "保溫杯", "送爸爸的禮物"]]
documents = [
    "Document: 修護滋養洗髮精 700ml | 深層滋養受損髮質 | 分類: 洗髮精",
    "Document: 不鏽鋼真空保溫杯 500ml | 長效保溫保冷 | 分類: 水壺保溫杯瓶",
    "Document: 純棉男士休閒襯衫 | 透氣舒適 | 分類: 男裝",
]

query_emb = model.encode(queries)
doc_emb = model.encode(documents)

print(model.similarity(query_emb, doc_emb))

Truncated Embeddings

model = SentenceTransformer("UNI-PIC/Retail-TW-Qwen3-Embedding-0.6B", truncate_dim=512)

Intended Use and Limitations

• Intended use: Semantic retrieval and ranking of retail product listings for Traditional Chinese shopping queries, typically as the dense component of a hybrid (dense and keyword) search system.
• Out-of-scope use: General-purpose text embedding outside the retail domain, and languages other than Traditional Chinese and English.
• Limitations: Performance depends on how closely a catalog resembles the training domain. Newly released brands, niche products, and listings with little descriptive text may be retrieved less reliably. Exact-match queries (brand or model number) should be handled by keyword search alongside dense retrieval. Evaluate on in-domain queries before deployment.

Citation

@misc{retailtwqwen3embedding06b,
  title     = {Retail-TW-Qwen3-Embedding-0.6B},
  author    = {President Information Corp. (PIC) and Uni-President Information Philippines Corp. (UNI-PIPC)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/UNI-PIC/Retail-TW-Qwen3-Embedding-0.6B}
}

Contact

• Taiwan: pic.net.tw/contactUs
• Philippines: pic.net.ph/about/contactUs