ModernBERT-TR Reranker

A 150M-parameter Turkish cross-encoder reranker to score (query, document) relevance.

• Base model: ytu-ce-cosmos/modernbert-tr-base.
• Distilled from Qwen/Qwen3-Reranker-8B.

Results

Reranking the top-100 of a first-stage retriever (ytu-ce-cosmos/modernbert-tr-embed) at max_seq=512. The uplift (Δ) is the reranker's contribution.

How was this model trained?

Question answering and counter argument distillation of Qwen3-Reranker-8B relevance scores into the 150M cross-encoder over Turkish question answering / information retrieval data using listwise KL.

Usage

transformers

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

tok = AutoTokenizer.from_pretrained("ytu-ce-cosmos/modernbert-tr-reranker")
model = AutoModelForSequenceClassification.from_pretrained("ytu-ce-cosmos/modernbert-tr-reranker").eval()

query = "Türkiye'nin başkenti neresidir?"
docs = ["Ankara, Türkiye'nin başkentidir.", "İstanbul en kalabalık şehirdir."]
enc = tok([query] * len(docs), docs, padding=True, truncation="longest_first",
          max_length=8192, return_tensors="pt")
with torch.no_grad():
    scores = model(**enc).logits.squeeze(-1)
ranking = sorted(zip(docs, scores.tolist()), key=lambda x: x[1], reverse=True)

sentence-transformers

from sentence_transformers import CrossEncoder
model = CrossEncoder("ytu-ce-cosmos/modernbert-tr-reranker")
scores = model.predict([(query, d) for d in docs])

ONNX Runtime

The onnx/ folder has the full graph, the output is the relevance logit:

import onnxruntime, numpy as np
sess = onnxruntime.InferenceSession("onnx/model.onnx")
feed = {k: v.numpy() for k, v in enc.items() if k in {i.name for i in sess.get_inputs()}}
logits = sess.run(None, feed)[0].squeeze(-1)

Text Embeddings Inference (TEI)

text-embeddings-router --model-id ytu-ce-cosmos/modernbert-tr-reranker --dtype float16
# POST /rerank  {"query": "soru", "texts": ["aday 1", "aday 2"]}

Training data

We used Turkish datasets msmarco-tr, squad-tr, fiqa-tr, nfcorpus-tr, quora-tr, scifact-tr for distillation by Qwen3-Reranker-8B, and Turkish counter-argument pairs from ArguAna machine-translated with TranslateGemma-27B. All training data was text-hash chceked against every MTEB(Turkish) test split.

Limitations

• Reported NDCG is rerank-of-top-100 over a first-stage retriever; absolute scores depend on that first stage.
• int8 ONNX reorders scores meaningfully lossy for a reranker; use fp32 for quality-sensitive ranking.
• Due to the lack of long form data in our training, the model's performance may degrade on long context input.

License & attribution

• License: apache-2.0.