amazon/poiss-bge-m3-reranker

POISS BGE-Reranker-v2-m3

BAAI/bge-reranker-v2-m3 fine-tuned on POISS to rerank point-of-interest candidates. It is a cross-encoder: query and candidate are encoded jointly, so unlike the bi-encoder it can exploit pairwise features like the distance between the user and the POI. Taking advantage of the cross-encoder's stronger expressive capabilities compared with a bi-encoder, the reranker was trained to incorporate the POI's quality score in addition to distance. Both distance and quality score are represented as natural-language buckets.

Intended position in the pipeline: retrieve with amazon/poiss-bge-m3-retriever, then rerank its top-k with this model. Output is a single logit per pair; higher is more relevant.

The numbers below were measured on the pool the paper uses — the retriever's top-1000 from the full 72M-POI corpus — which is not distributed with the dataset. Reranking the dataset's candidates column instead is an easier setting (every graded positive is present by construction) and gives different, non-comparable numbers. To reproduce this table, rebuild the pool by embedding the corpus with the retriever above.

Results

Reranking the fine-tuned retriever's top-1000 on the POISS test split (42,018 queries, percentages):

System	P@5	P@20	MRR	N@5	N@20
BGE-M3 retriever (bi-encoder order)	57.6	35.5	85.5	61.5	62.6
BGE-Reranker-v2-m3 zero-shot	24.6	14.3	50.0	32.1	34.8
this model	63.0	38.3	89.7	66.8	66.6

Per-intent nDCG@20: Search 76.8, Detail 78.9, Recommend 56.3, Things-to-do 55.0. Precision is higher on open-ended intents simply because their relevant sets are larger; the rank-aware metrics show the ordering there is the harder problem.

Usage

The model scores a text pair: a query string and a POI string. Field order, prompt labels, and the bucket vocabularies below are part of the trained format.

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "amazon/poiss-bge-m3-reranker"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).eval()

SCORE_BUCKETS = [  # upper bound (exclusive), label
    (1.0, "Very Poor"), (1.5, "Poor"), (2.0, "Somewhat Poor"), (2.5, "Fair"),
    (3.0, "Average"), (3.5, "Good"), (4.0, "Very Good"), (4.5, "Excellent"),
    (5.0, "Outstanding"), (5.1, "Perfect"),
]
DISTANCE_BUCKETS = [  # upper bound in km (exclusive), label
    (2.0, "extremely close"), (4.0, "very close"), (6.0, "close"), (8.0, "nearby"),
    (10.0, "moderately close"), (15.0, "moderately far"), (20.0, "far"),
    (30.0, "quite far"), (50.0, "very far"), (320.0, "extremely far"),
]


def bucket(value, buckets):
    for upper, label in buckets:
        if value < upper:
            return label
    return buckets[-1][1]


def format_query(text):
    return f"Query: {text[:100]}"


def format_poi(name, categories, distance_km, score, address):
    # categories: a list; the first 10 are kept and joined by a space
    fields = [f"Name: {name[:60]}", f"Categories: {' '.join(categories[:10])}"]
    if distance_km is not None and distance_km >= 0:
        fields.append(f"Distance: {bucket(distance_km, DISTANCE_BUCKETS)}")
    fields.append(
        f"Score: {bucket(score, SCORE_BUCKETS)}" if score and score > 0 else "Score: None"
    )
    fields.append(f"Address: {address[:60]}")
    return "; ".join(fields)


query = format_query("pizzeria near the colosseum")
pois = [
    format_poi("Pizzeria Da Enzo", ["pizza_restaurant", "restaurant"], 0.6, 4.3,
               "Via dei Fori Imperiali 1, Rome"),
    format_poi("Oslo Dental Clinic", ["dentist"], 2100.0, 3.1, "Storgata 10, Oslo"),
]

batch = tokenizer([query] * len(pois), pois, padding=True, truncation=True,
                  max_length=256, return_tensors="pt")
with torch.no_grad():
    scores = model(**batch).logits.squeeze(-1)
print(scores)  # higher = more relevant

Details that matter:

• Field order is fixed: Name, Categories, Distance, Score, Address, joined by "; ".
• Categories is the first 10 category labels joined by a space; the query is truncated at 100 characters, and anything after a "; Position:" marker is stripped from it.
• Score is the POISS 0–5 quality score, shipped with the dataset as data/poi_quality_score.parquet (it is not an Overture field). When it is missing or non-positive the field is still emitted, as Score: None. Values above 5 are halved and clamped.
• The score is sparse: it covers 30.7% of query–candidate pairs in POISS (43.7% of the graded positives). Score: None is therefore the common case, and is what this model saw for about two thirds of its training pairs — not a degraded mode.
• Distance is the haversine distance in km between the user coordinate and the POI, and is omitted entirely when unavailable.
• Truncation: 256 tokens for the pair; the source cutoffs are 100 characters for the query, 60 for the name and address, 10 categories.

Training

Fine-tuned from BAAI/bge-reranker-v2-m3 on the POISS train split (254,089 training and 13,374 validation records), with POI content resolved from Overture Places release 2026-03-18.

Objective	binary cross-entropy on the pair logit
Hard negatives	100 mined per query, 16 sampled per step
Optimizer	AdamW, lr 1e-6, gradient clipping 1.0
Batch size	4 per GPU × 8 GPUs
Max sequence length	256
Precision / hardware	mixed precision, 8× NVIDIA A100 40GB
Schedule	up to 20 epochs, best checkpoint by validation nDCG@20
Selected checkpoint	epoch 18, validation nDCG@20 0.792

Negatives are the candidates the LLM judge marked non-relevant, padded from the reranked top-300 starting at rank 30. Positives are the graded relevant POIs.

The quality score the model consumes is LLM-generated from Overture POI metadata.

Reproducibility note

The published weights are the checkpoint used for the paper, re-expressed in the stock XLMRobertaForSequenceClassification layout; all 393 tensors are bit-identical to the training checkpoint.

Intended use and limitations

Intended for research on POI search reranking. Reranking quality is bounded by the candidate set: relevant POIs the retriever never surfaces cannot be recovered, and on POISS the fine-tuned retriever's R@1k is 95.6, so a few percent of graded POIs are out of reach by construction. Coverage is limited to 7 languages concentrated in the Americas and Europe, so behaviour outside those locales is untested.

Citation

@inproceedings{maritan-etal-2026-poiss,
    title = "{POISS}: A Large-Scale Multilingual Dataset for Point-of-Interest Search",
    author = "Maritan, Nicola  and
      Moschitti, Alessandro  and
      Borazio, Federico  and
      Zhou, Xiaokun  and
      Bai, Zhengwei",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    note = "To appear",
}

License

These weights are a derivative work of BAAI/bge-reranker-v2-m3, distributed by BAAI under the Apache License 2.0, and are released under the same license. A copy is included as LICENSE, and NOTICE records the required statement of modifications: all weights were modified by supervised fine-tuning on the POISS train split with a binary cross-entropy objective, with no change to architecture, configuration, vocabulary or tokenizer.

The base model itself derives from BAAI/bge-m3 (MIT), which derives from XLM-RoBERTa (MIT).

The training data, POISS, is Apache-2.0 and places no additional restriction on these weights.