cp500/multilingual-automotive-sparse

Multilingual Automotive Sparse Retrieval Corpus

A synthetic training corpus for fine-tuning multilingual sparse retrieval models (SPLADE-family) on automotive, supply-chain, and geopolitics content. Designed to teach a model cross-lingual alignment: a Japanese query should retrieve the right English passage, and vice versa.

Corpus contents

File	Rows	Purpose
concepts.jsonl	4995	Raw concept records
train_triplets.jsonl	134880	Flattened (query, positive, negative) triplets
eval_pairs.jsonl	4491	Held-out 10% for retrieval evaluation

Concept record schema (concepts.jsonl)

{
  "query":          {"en": str, "ja": str, "ko": str},
  "positive":       {"en": str, "ja": str, "ko": str},
  "hard_negatives": [{"en": str, "ja": str, "ko": str}, × 5],
  "easy_negative":  {"en": str},
  "_meta": {
    "concept_id": str,
    "category": str,
    "tags": [str],
    "subject": str, "predicate": str, "object": str,
    "model": "us.anthropic.claude-haiku-4-5-20251001-v1:0",
    "usage": {"input_tokens": int, "output_tokens": int, ...}
  }
}

Training triplet schema (train_triplets.jsonl)

{
  "query": str,
  "positive": str,
  "negative": str,
  "query_lang": "en"|"ja"|"ko",
  "positive_lang": "en"|"ja"|"ko",
  "negative_lang": "en"|"ja"|"ko",
  "concept_id": str,
  "category": str,
  "negative_kind": "hard"|"easy",
  "negative_rank": int
}

Each concept fans out into up to 30 training triplets covering all 9 query×positive language-pair combinations, with hard negatives preferentially matching the query language.

Category distribution

category	count
energy_commodities	850
pharma_biotech	843
semiconductors_hardware	842
geopolitics_defense	840
automotive_mobility	828
finance_capital_markets	792

Generation method

Records were synthesized via Anthropic Claude Haiku 4.5 on AWS Bedrock, using a deterministic stratified sample of 2500 concept seeds drawn from 8 automotive / supply-chain / geopolitics scenarios. The model was prompted to produce faithful translations across EN/JA/KO with native hard-negative phrasing (not translated-from-English).

Hard negative taxonomy

Each concept has 5 hard negatives, deliberately varied:

1. Same entity / different action
2. Similar action / different entity
3. Same region / unrelated industry
4. Similar entity + topic / different polarity or timeframe
5. Adjacent domain

Usage

from datasets import load_dataset

# Full corpus
ds = load_dataset("cp500/multilingual-automotive-sparse")

# Training triplets only
triplets = load_dataset("cp500/multilingual-automotive-sparse", data_files="train_triplets.jsonl")

# Held-out eval pairs
eval_ = load_dataset("cp500/multilingual-automotive-sparse", data_files="eval_pairs.jsonl")

Intended use

• Fine-tuning multilingual SPLADE / sparse retrieval models (MarginMSE, contrastive loss, in-batch negatives).
• Cross-lingual retrieval research: measure whether models learn lexical-vocab-space alignment across scripts.
• FLOPS-regularization experiments in the automotive-intelligence domain.

Limitations

• All passages are synthetic — companies, dollar amounts, and events are plausible but fictional. Not suitable for factual question answering.
• Machine-generated CJK may have occasional register issues — native-speaker review has not been performed.
• Domain is narrow (automotive / supply chain / geopolitics). Performance on out-of-domain text will require additional corpora.

License

Apache 2.0. Generated with Anthropic Claude — users should review Anthropic's Acceptable Use Policy for downstream applications.

Citation

@misc{cp500-multilingual-automotive-sparse-2026,
  author = {Charles P},
  title  = {Multilingual Automotive Sparse Retrieval Corpus},
  year   = {2026},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/datasets/cp500/multilingual-automotive-sparse}
}