japanese-reranker-small-v2-coreml

English | 日本語

Model Summary

This is an unofficial Core ML conversion of hotchpotch/japanese-reranker-small-v2 (a lightweight Japanese reranker based on ModernBERT), for running directly on iOS/macOS. All credit for the original model goes to its author.

• Output: a relevance score in [0, 1] (sigmoid-applied logit). The query/document pair is tokenized together as a single sequence.

• Models are split by fixed sequence length.

• Live demo: Ruri Atlas — explore Japanese & English sentence embeddings as a 3-D map in your browser (WebGPU), switch embedding models, toggle the reranker, and see where a prompt-injection lands. Nothing is sent to a server.

File list

File	Seq len	Precision	Approx. size
japanese-reranker-small-v2_seq128_fp16.mlpackage	128	fp16	~280MB
japanese-reranker-small-v2_seq256_fp16.mlpackage	256	fp16	~280MB
japanese-reranker-small-v2_seq512_fp16.mlpackage	512	fp16	~280MB
japanese-reranker-small-v2_seq128_int8.mlpackage	128	int8	~140MB
japanese-reranker-small-v2_seq256_int8.mlpackage	256	int8	~140MB
japanese-reranker-small-v2_seq512_int8.mlpackage	512	int8	~140MB

Swift (Core ML) usage example

import CoreML

let configuration = MLModelConfiguration()
configuration.computeUnits = .all
let model = try MLModel(
    contentsOf: Bundle.main.url(forResource: "japanese-reranker-small-v2_seq256_int8", withExtension: "mlmodelc")!,
    configuration: configuration
)

let inputIDs: MLMultiArray = ...       // shape [1, 256], Int32 (query+document tokenized as a pair)
let attentionMask: MLMultiArray = ...  // shape [1, 256], Int32

let input = try MLDictionaryFeatureProvider(dictionary: [
    "input_ids": MLFeatureValue(multiArray: inputIDs),
    "attention_mask": MLFeatureValue(multiArray: attentionMask),
])
let output = try model.prediction(from: input)
let score = output.featureValue(for: "relevance_score")!.multiArrayValue![0]
// relevance score in [0, 1]

Python verification example

import numpy as np
import coremltools as ct
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("hotchpotch/japanese-reranker-small-v2")
mlmodel = ct.models.MLModel("japanese-reranker-small-v2_seq256_int8.mlpackage")

query = "瑠璃色はどんな色？"
document = "瑠璃色（るりいろ）は、紫みを帯びた濃い青。"
enc = tokenizer(query, document, return_tensors="np", padding="max_length", truncation=True, max_length=256)

out = mlmodel.predict({
    "input_ids": enc["input_ids"].astype(np.int32),
    "attention_mask": enc["attention_mask"].astype(np.int32),
})
print(out["relevance_score"])

Notes

• This is a community conversion, not an official release from the original author.

Security

Audited against its upstream with model-audit-lite: weight format, bundled code, and a machine-readable lineage (ML-BOM). Details, checksums and how to reproduce: SECURITY.md.

---

モデルの概要

hotchpotch/japanese-reranker-small-v2（軽量な日本語reranker、ModernBERTベース）を、iOS/macOS (Core ML) で直接動かせるように変換したものです。元モデルの著作権はその作者に帰属します。

• 出力: 0〜1の関連度スコア（sigmoid適用済みのlogit）。query/documentのペアを1つのシーケンスとしてまとめてトークナイズして入力します。

• 系列長ごとに固定長でモデルを分けています。

• ライブデモ: Ruri Atlas — 日本語・英語の文の埋め込みを、ブラウザ上（WebGPU）で3Dの地図として見られます。埋め込みモデルの切り替え、リランカーのON/OFF、プロンプトインジェクションの効き方の可視化もできます。文章はサーバーに送られません。

ファイル一覧

ファイル	系列長	精度	サイズ目安
japanese-reranker-small-v2_seq128_fp16.mlpackage	128	fp16	約280MB
japanese-reranker-small-v2_seq256_fp16.mlpackage	256	fp16	約280MB
japanese-reranker-small-v2_seq512_fp16.mlpackage	512	fp16	約280MB
japanese-reranker-small-v2_seq128_int8.mlpackage	128	int8	約140MB
japanese-reranker-small-v2_seq256_int8.mlpackage	256	int8	約140MB
japanese-reranker-small-v2_seq512_int8.mlpackage	512	int8	約140MB

Swift (Core ML) での使用例

import CoreML

let configuration = MLModelConfiguration()
configuration.computeUnits = .all
let model = try MLModel(
    contentsOf: Bundle.main.url(forResource: "japanese-reranker-small-v2_seq256_int8", withExtension: "mlmodelc")!,
    configuration: configuration
)

let inputIDs: MLMultiArray = ...       // shape [1, 256], Int32（query+documentをペアでトークナイズ）
let attentionMask: MLMultiArray = ...  // shape [1, 256], Int32

let input = try MLDictionaryFeatureProvider(dictionary: [
    "input_ids": MLFeatureValue(multiArray: inputIDs),
    "attention_mask": MLFeatureValue(multiArray: attentionMask),
])
let output = try model.prediction(from: input)
let score = output.featureValue(for: "relevance_score")!.multiArrayValue![0]
// 0〜1の関連度スコア

Python での検証例

import numpy as np
import coremltools as ct
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("hotchpotch/japanese-reranker-small-v2")
mlmodel = ct.models.MLModel("japanese-reranker-small-v2_seq256_int8.mlpackage")

query = "瑠璃色はどんな色？"
document = "瑠璃色（るりいろ）は、紫みを帯びた濃い青。"
enc = tokenizer(query, document, return_tensors="np", padding="max_length", truncation=True, max_length=256)

out = mlmodel.predict({
    "input_ids": enc["input_ids"].astype(np.int32),
    "attention_mask": enc["attention_mask"].astype(np.int32),
})
print(out["relevance_score"])

備考

• 本変換は非公式のコミュニティ版です。

セキュリティー

model-audit-lite で変換元と突き合わせて監査済みです（重みの形式、同梱コード、機械可読な系譜＝ML-BOM）。詳細・チェックサム・再現方法は SECURITY.md をご覧ください。