TY-ecomm-asure-relevance

A multilingual cross-encoder that scores how relevant a product is to a query, over the four ESCI classes. Distilled from Trendyol/Trendyol-LLM-Asure-12B.

Model Details

Developed by	Trendyol - LLM & Core NLP Team
Model type	Cross-encoder sequence-pair classifier (4 classes)
Base model	Trendyol/TY-ecomm-embed-multilingual-base-v1.2.0
Parameters	311.3M
Languages	Turkish, Arabic, Romanian, English
Max sequence length	512 tokens
Inputs	input_ids, attention_mask (no token_type_ids)
Output	4 raw logits, apply softmax yourself
License	Apache-2.0

Labels

id	label	meaning	example for query iphone 15
0	irrelevant	no meaningful relation	Samsung Galaxy S24 case
1	complementary	related, serves a different need	iPhone 15 screen protector
2	substitute	a reasonable alternative	Samsung Galaxy S24 blue
3	relevant	directly satisfies the query intent	iPhone 15 128 GB blue

Intended Use

Use this model to classify product candidates for a user query by estimating whether each product is irrelevant, complementary, substitute, or relevant.

Limitations

The model can make incorrect relevance judgments and only uses the text fields provided at inference time. It does not see product images and should not be treated as having general world knowledge beyond the query-product text pair.

Input format

The first sequence is the query. The second is the product document: seven fields, one per line, separated by \n, using these exact English field names. They are the product attributes the model saw in training, so keep the names, the order and the separators.

title: {title}
category: {category}
brand: {brand}
gender: {gender}
age_group: {age_group}
marketplace_sellers: {seller1 | seller2 | ...}
attributes: {key1}: {value1}, {key2}: {value2}, ...

rule	value
missing field	the literal string unknown
sellers	deduplicated case-insensitively, first 10, joined with |
attributes	drop any whose value exceeds 100 characters, then keep the first 30
attributes rendering	key: value, joined with , on one line; attributes: unknown if none
text normalization	lowercase, NFC, and i̇ to i

The tokenizer encodes the pair as <s> query </s></s> document </s>.

Sample input and output

Input:

query: "dönen mumluk"

title: truro mumluk naturel (11x14,5x5 cm)
category: ev ve mobilya / ev dekorasyon / şamdan mumluk
brand: bella maison
gender: unknown
age_group: unknown
marketplace_sellers: bella maison
attributes: renk: turuncu, materyal: ahşap, parça sayısı: 1, color detail: şeffaf

Output:

{'irrelevant': 0.0368, 'complementary': 0.0129, 'substitute': 0.9253, 'relevant': 0.025}

A rotating candle holder was searched for and a plain wooden one was found, so the product scores as substitute rather than relevant.

Installation

pip install torch "transformers>=4.57,<5"

transformers 5.x is not supported yet.

How to use

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "Trendyol/TY-ecomm-asure-relevance"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(model_id, trust_remote_code=True).eval()

query = "dönen mumluk"
document = "\n".join([
    "title: truro mumluk naturel (11x14,5x5 cm)",
    "category: ev ve mobilya / ev dekorasyon / şamdan mumluk",
    "brand: bella maison",
    "gender: unknown",
    "age_group: unknown",
    "marketplace_sellers: bella maison",
    "attributes: renk: turuncu, materyal: ahşap, parça sayısı: 1, color detail: şeffaf",
])

enc = tok(query, document, truncation=True, max_length=512, return_tensors="pt")
with torch.no_grad():
    probs = model(**enc).logits.softmax(-1)[0]
print({model.config.id2label[i]: round(p.item(), 4) for i, p in enumerate(probs)})

trust_remote_code=True is required: both the GTE encoder and the classification head are defined in this repository rather than in transformers.

Scoring many candidates for one query

queries = [query] * len(documents)
enc = tok(queries, documents, truncation=True, max_length=512,
          padding=True, return_tensors="pt")
with torch.no_grad():
    scores = model(**enc).logits.softmax(-1)[:, 3]  # P(relevant)

Training data

~1.5M query-product pairs in Turkish, Arabic, Romanian, English languages.

License

Apache-2.0

Citation

@misc{trendyol-ty-ecomm-asure-relevance,
  title  = {TY-ecomm-asure-relevance: Trendyol E-commerce Multilingual Query-Product Relevance Cross-Encoder},
  author = {Trendyol - LLM & Core NLP Team},
  year   = {2026},
  url    = {https://huggingface.co/Trendyol/TY-ecomm-asure-relevance}
}

Model Card Authors

Trendyol - LLM & Core NLP Team