KazeJiang/CLICKER-bge-m3-retriever

CLICKER bge-m3 Retriever

This repository provides the fine-tuned BAAI/bge-m3 retriever used in CLICKER: Cross-Lingual Knowledge Editing via In-Context Learning with Adaptive Stepwise Reasoning.

The model is the Step-1 relevance-aware dense retriever in CLICKER. It is used to retrieve edited knowledge from an edit base and decide whether the retrieved fact is relevant enough to inject into the in-context knowledge editing prompt. It is not a standalone generative language model and does not include the full CLICKER pipeline.

Model Details

• Base model: BAAI/bge-m3
• Model type: SentenceTransformer / dense text encoder
• Backbone architecture: XLM-RoBERTa
• Embedding size: 1024
• Maximum sequence length: 8192 tokens
• Pooling: CLS pooling followed by normalization
• Similarity: cosine similarity
• Training objective: triplet loss
• Training data: triplets constructed from the Multi-CounterFact training set
• Languages used in the paper: English, German, French, Japanese, Chinese

Intended Use

This model is intended for relevance-aware retrieval in cross-lingual dynamic knowledge editing, especially as the retriever component of CLICKER.

In CLICKER, facts in an edit base are encoded with this model and indexed with FAISS. Given a user query, the retriever returns the nearest edited fact and a cosine similarity score. A tuned threshold is then used to decide whether the fact should be injected into the downstream in-context editing prompt.

Usage

Install dependencies:

pip install -U sentence-transformers faiss-cpu

Load the model:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("KazeJiang/CLICKER-bge-m3-retriever")

texts = [
    "What is the official language of the United Nations?",
    "The official language of the United Nations is Indonesian.",
]

embeddings = model.encode(texts, normalize_embeddings=True)
similarity = embeddings @ embeddings.T
print(similarity)

Minimal retrieval example:

import faiss
import numpy as np
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("KazeJiang/CLICKER-bge-m3-retriever")

edit_facts = [
    "The official language of the United Nations is Indonesian.",
    "Danielle Darrieux recorded for Capitol Nashville.",
]

fact_embeddings = model.encode(edit_facts, normalize_embeddings=True)
index = faiss.IndexFlatIP(fact_embeddings.shape[1])
index.add(np.asarray(fact_embeddings, dtype="float32"))

query = "Which language is used by the United Nations?"
query_embedding = model.encode([query], normalize_embeddings=True)

scores, indices = index.search(np.asarray(query_embedding, dtype="float32"), k=1)
print(edit_facts[indices[0][0]], float(scores[0][0]))

For the full CLICKER pipeline, please refer to the project code repository.

Note on the ReMaKE Baseline

The ReMaKE baseline discussed in the CLICKER paper also uses a retriever component. This repository only releases the fine-tuned BAAI/bge-m3 retriever used by CLICKER. If you want to reproduce the ReMaKE retriever exactly, please request the corresponding retriever checkpoint from the original ReMaKE authors.

Training

The retriever was fine-tuned from BAAI/bge-m3 using triplet training examples from Multi-CounterFact. Each triplet contains a query, a preferred candidate, and a less-preferred candidate. Positive triplets encourage cross-lingual matching between related queries and edited facts, while negative triplets teach the model to prefer [NULL] over irrelevant edited facts.

The model was trained with triplet loss and margin 0.1, following the setup described in the paper.

Limitations

This model is specialized for relevance-aware retrieval in the CLICKER setting. It should not be treated as a general-purpose factuality model or as a generative model. The retrieval threshold should be tuned on a validation set for the target language pair and edit-base distribution.

Citation

If you use this model, please cite:

@inproceedings{jiang-etal-2026-clicker,
  title = {{CLICKER}: Cross-Lingual Knowledge Editing via In-Context Learning with Adaptive Stepwise Reasoning},
  author = {Jiang, Zehui and Zhao, Xin and Kumadaki, Yuta and Yoshinaga, Naoki},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026}
}