---
title: 'AdamDB: Multilingual Biographical KB'
url: https://www.emergentmind.com/topics/adamdb
type: topic
---

# AdamDB: Multilingual Biographical KB

AdamDB is the foundational component of the ADAM framework, designed as an extensive, structured, multilingual, and multimodal biographical knowledge base for evaluating and improving large language models in biographical reasoning. It covers over 4 million individuals across geography, time, and profession; spans 595 languages; includes both textual biographical records and associated face images; and serves as the substrate for both AdamBench, a cognitively structured evaluation suite, and AdamRAG, a retrieval-augmented generation system tailored to biography [2509.22991].

## 1. Scope, coverage, and corpus composition

AdamDB is described as covering over 4 million unique individuals across history, with birth dates quantized to 50-year bins for cluster sampling. Its geographic scope is global, representing individuals from 200+ countries, with at least 10 individuals per country or territory to ensure worldwide inclusion. The distribution by continent is reported as Europe 52.5%, North America 27.4%, Asia 8.4%, South America 6.0%, Oceania 3.9%, and Africa 1.9% [2509.22991].

The database is multilingual at unusual scale. Biographical records span 595 languages. Reported language coverage includes English with 3,973,119 names (99.7%), Dutch with 3,423,044 (85.9%), Spanish with 2,593,617 (65.1%), French with 2,070,396 (52.0%), German with 1,921,806 (48.2%), Italian with 1,580,213 (39.7%), Portuguese with 1,225,397 (30.8%), and Polish with 649,224 (16.3%). The principal modalities are structured and unstructured textual data—names, biographies, dates, nationalities—and associated face images when available or crawled [2509.22991].

A comparative summary reported in the source material situates the resource against prior biography datasets:

| System | Records | Languages | Countries |
|---|---:|---:|---|
| ADAM | 4,016,647 | 595 | Global |
| BiographySampo | 13,100 | 1 | 1 |
| BiographyNet | 125,000 | 1 | 1 |
| Networked Pantheon | 11,341 | Limited | Global |

This scale is significant because AdamDB is explicitly intended to reduce the English, popularity, and modality biases that characterize predecessor datasets, while retaining sufficient coverage of lesser-known figures for factuality analysis and retrieval-grounded inference [2509.22991].

## 2. Collection pipeline, filtering, and record alignment

The construction pipeline begins from the WikiDBS dataset, described as large-scale relational tables containing Wikipedia and Wikidata extractions. Filtering then proceeds in two stages. In Stage 1, foreign key columns are heuristically matched with person-related terms such as “surname.” In Stage 2, Named Entity Recognition is used to retain only columns strongly associated with PERSON entities [2509.22991].

Biographical records are extracted by row, merged through name-based heuristics, and validated with further NER. De-duplication and alignment are performed against Wikidata Q-Ids. Duplicate entries are consolidated by choosing modal values for biography, nationality, birth date, and birthplace. Multilinguality is expanded by retrieving and linking all available language variants for names from Wikidata [2509.22991].

Coverage control is strict. Individuals are retained only if they have valid, non-null biography, birth date, nationality, and birthplace. Popularity is incorporated through annual page views from English Wikipedia (2024); individuals with zero page views are discarded, but minimum country-wise representation is enforced. For multimodal support, each individual is linked either to known Wikipedia face images or to crawled image URLs, with copyright compliance explicitly noted [2509.22991].

These steps define AdamDB not merely as a scraped corpus but as an aligned biographical database with identifier resolution, record consolidation, multilingual linkage, and controlled inclusion criteria. A plausible implication is that the system is optimized not only for scale but also for downstream retrieval fidelity, disambiguation, and stratified benchmarking.

## 3. Design goals and representational role

AdamDB was created to enhance LLM training and evaluation in biographical reasoning beyond rote memorization and into higher-order cognitive skills such as analysis and synthesis. It is also intended to cover global, multilingual, and multimodal biographical knowledge; reduce factual hallucination through structured, validated factual data retrievable at inference time; support cognitively rigorous benchmarking through AdamBench; and serve as a resource for retrieval-augmented generation systems requiring robust, accurate, and diverse biographical ground truth [2509.22991].

The representational role of AdamDB is easy to misstate. AdamDB is the database; AdamBench is the evaluation suite drawing questions from it; and AdamRAG is the retrieval system that grounds model outputs in its entries. This distinction matters because the paper treats the three components as an integrated framework rather than interchangeable names. Questions in AdamBench are sampled from AdamDB, while AdamRAG retrieves relevant AdamDB context at inference time [2509.22991].

The database is also designed around factual integrity and hallucination mitigation. The source emphasizes deduplication via Wikidata Q-Ids, strict field requirements, page-view-based popularity stratification, multilingual labels, and associated face images. These mechanisms are presented as enhancing accuracy and reproducibility, especially when AdamDB is paired with AdamRAG [2509.22991].

## 4. AdamBench: cognitively structured evaluation derived from AdamDB

AdamBench is the evaluation suite built from AdamDB. Its role is to assess LLM biographical reasoning through multiple-choice questions presented in both English and the individual’s native language. The benchmark is organized around Bloom’s Taxonomy across six cognitive levels: Remembering, Understanding, Applying, Analyzing, Evaluating, and Creating. Questions may also be multimodal, for example by including an image alongside a name [2509.22991].

Sampling is proportional and stratified by country and popularity. The base cluster count for each country is defined by

$$
k = \lceil (\text{country population proportion} \times 5) + 0.01 \rceil
$$

with a minimum of one per country enforced. Individuals are stratified into high, medium, and low popularity groups. Within this structure, $\text{k-means}$ clustering is performed on concatenated feature vectors consisting of quantized birth date and a normalized BERT embedding of the biography, with the birth date feature weighted for balance. Cluster centroids are selected, or the most popular instance is selected for low-popularity clusters, in order to maximize diversity [2509.22991].

The resulting benchmark samples about 1,650 individuals. Evaluation is implemented through EleutherAI lm-eval-harness for automation and the Khayyam Challenge platform for multilingual and multimodal support. The benchmark therefore operationalizes AdamDB’s coverage into a cognitively stratified testbed rather than a simple factual lookup task [2509.22991].

## 5. AdamRAG and factual grounding in biography

AdamRAG is the retrieval-augmented generation subsystem tailored to biography. Given a user query, relevant biographical entries are retrieved from AdamDB, and the retrieved context augments the LLM input prompt to support factual answering. The system supports both text-based and multimodal queries, including those involving face images [2509.22991].

Disambiguation is handled differently by modality. For text queries, AdamRAG uses LaBSE embeddings together with nationality and birth date filters. For image-based queries, face embeddings are matched against AdamDB and then filtered for plausible biographical fits. Popularity-aware retrieval incorporates Wikipedia page-view metrics to favor high-relevance information [2509.22991].

The paper frames AdamRAG as the primary mechanism for addressing factual hallucination, especially for lesser-known individuals. The stated strategy is structured grounding: LLM responses can be made to rely on verified AdamDB entries. The popularity-sensitive design ensures that less known individuals are not ignored, and multilingual plus cross-modality support reduces errors caused by language gaps or ambiguous queries. Empirically, AdamRAG is reported to substantially reduce hallucination and improve accuracy, especially on lesser-known figures; the largest gains occur on lower-order reasoning, while multimodal input via face images offers smaller, less consistent improvements than retrieval [2509.22991].

## 6. Empirical behavior, limitations, and naming ambiguity

The primary evaluation metric is accuracy in multiple-choice settings, analyzed by Bloom level, language, popularity, and modality. Reported results indicate that closed-source models such as GPT-4 and Gemini Flash achieve 85–95%+ accuracy, with Gemini Flash~2.5 frequently exceeding 95% when using retrieval. Open-source models such as Gemma3-12b-it and Qwen2.5-7b have substantially lower baselines, below 60%, but AdamRAG raises accuracy markedly, up to about 80% on mid- and high-popularity individuals. Retrieval is said to close the gap between open- and closed-source models, with the largest effect on lower cognitive levels and on lesser-known individuals. Popularity strongly mediates accuracy, retrieval removes most language-based performance differences for factual lower-Bloom questions, and image inclusion yields smaller, mixed improvements relative to retrieval augmentation [2509.22991].

Data availability and ethics are addressed directly. AdamDB and AdamBench are to be publicly released via Hugging Face. Raw data respect copyright and privacy because images are not redistributed; they are only referenced via public URLs. This design separates the knowledge base from direct redistribution of media assets [2509.22991].

The name “AdamDB” is not unique to this framework. In asteroid reconstruction, the ADAM paper describes a role for asteroid model databases in which ADAM-based models can be systematically produced and catalogued, with rapid shape and spin solutions for hundreds or thousands of asteroids and standardized estimation of model reliability [1501.05958]. In Alzheimer’s disease analysis, “AdamDB” is used for a knowledge-centric interpretative AI that integrates microbiome profiles, clinical datasets, and a literature knowledge base of 80,909 unique AD-related publications embedded into over 2 million segments [2501.08324]. This suggests that “AdamDB” is a domain-dependent label rather than a single standardized resource. In the biographical-reasoning literature, however, the term denotes the multilingual and multimodal biographical database within the ADAM framework [2509.22991].

Source: https://www.emergentmind.com/topics/adamdb