Cross-lingual Biography Enrichment via Claim Extraction and Alignment

• The study addresses uneven multilingual Wikipedia coverage, where English is often treated as the default source despite non-English editions containing richer locally relevant information for long-tail figures.Prior work covered structured-data biography generation or retrieval-augmented full biographies, but did not directly study enriching existing English biographies with curated non-English Wikipedia evidence.Source: §1 · p. 1
• Both translation-based and claim-based generation produce more supported additions than raw cross-lingual generation, while claim-based generation substantially lowers hallucination.The comparison uses the same rewriting generators and varies only how the non-English biography is represented.Source: Figure 1 · p. 1, §1 · p. 2
• The task enriches an existing English biography of a woman with information present in a corresponding non-English biography but absent from the English article.This differs from generating biographies from scratch because it starts with an existing English article and uses a curated non-English article as additional evidence.Source: §1 · pp. 1–2
• The claim-based framework extracts both biographies into a shared English claim space, aligns claim pairs, filters already-covered information, and supplies the remaining enrichment claims for controlled rewriting.This makes evidence selection explicit, reduces long-context and implicit cross-lingual alignment demands, and supports intermediate evaluation.Source: §1 · p. 2, Figure 2 · p. 2
• The work provides CLAW-4L benchmarks for French-, Chinese-, and Azerbaijani-to-English enrichment, cross-lingual claim extraction, and English claim-pair relation classification.The benchmarks are named CLAW-4L, CLAW-4L-CX, and CLAW-4L-RC.Source: §1 · p. 2

Wikipedia Biography Generation and Enrichment

• Biography-generation systems evolved from short infobox- or Wikidata-based outputs toward retrieved, full-length biographies, but existing approaches do not target cross-lingual enrichment of full biographies.Structured-data systems include Lebret et al. (2016), Chisholm et al. (2017), Vougiouklis et al. (2018), Kaffee et al. (2018), and Mille et al. (2024); retrieval-grounded work includes Banerjee and Mitra (2016), Zhang et al. (2025), Fan and Gardent (2022), and Adak et al. (2025).Source: §2 · pp. 2–3

Claim Extraction, Decomposition, and Alignment

• Claim-extraction research provides candidate frameworks for decomposing and verifying text, but is primarily English-focused and depends on task-specific schemas or language-specific tools.The work evaluates cross-lingual adaptations of open information extraction, fact extraction, and LLM-based methods including FactScore (Min et al., 2023), FactCheck-GPT (Wang et al., 2024), VeriScore (Song et al., 2024), DnDScore (Wanner et al., 2025), and Claimify.Source: §2 · p. 3

Cross-lingual Fact-checking and Claim Alignment

• Cross-lingual fact-checking studies multilingual claim detection, retrieval, and verification, whereas this setting extracts English claims from non-English biographies and aligns them for enrichment.This distinction is made relative to Gupta and Srikumar (2021) and Chang et al. (2023).Source: §2 · p. 3
• CLAW-4L-RC uses a finer-grained relation space that distinguishes exact coverage, specificity differences, complementarity, contradiction, and unrelatedness beyond binary or three-way NLI labels.The relation space supports enrichment-oriented claim selection, with NLI and fact-verification models serving as baselines (Bowman et al., 2015; Tang et al., 2024).Source: §2 · p. 3

Task and Method

• Given paired biographies of the same person, the output preserves the English biography’s content while incorporating information supported by its French, Chinese, or Azerbaijani counterpart.The input pair consists of an English biography and a biography in one of the three specified languages.Source: §3 · p. 3
• The method rewrites the original English biography with additional English claims extracted from the non-English biography but absent from the English source.Claims are defined as predicates with arguments and modifiers, optionally including hedges; in these factual biographies, they largely correspond to facts.Source: §3 · p. 3

4 Datasets

• The three benchmarks respectively evaluate biography enrichment, claim extraction, and relation classification within the proposed framework.CLAW-4L links English biographies with French, Chinese, or Azerbaijani counterparts; CLAW-4L-CX evaluates claim extraction models, while CLAW-4L-RC evaluates relation classifiers.Source: §4 · p. 3

4.2 CLAW-4L-CX

• CLAW-4L-CX contains 600 sentences manually annotated with English claims, balanced across English and non-English languages and claim-complexity levels.The benchmark includes 300 English and 300 non-English sentences, with human annotators verifying, revising, and supplementing automatically generated claim drafts; Aligned-F1 exceeds 91% for every language.Source: §4 · p. 3, §4.2 · p. 4

4.3 CLAW-4L-RC

• CLAW-4L-RC contains 600 claim pairs labeled as exact alignment, partial alignment or enrichment, contradiction, or irrelevance.The final distribution is 100 exact-alignment, 300 partial-alignment/enrichment, 100 contradicted, and 100 not-relevant pairs, constructed through validation and controlled perturbation or rewriting.Source: §4 · p. 3, §4.3 · pp. 4–5

4.1 CLAW-4L

• Non-English biographies contain 3.4× more tokens, 3.4× more sentences, and 2.9× more claims on average than their English counterparts.CLAW-4L comprises 300 paired biographies, with 100 biographies per non-English language and occupation-aware balancing; language-calibrated ratios are used only as heuristic proxies for richness.Source: §4.1 · p. 4, Table 1 · p. 5

5.1 Classifying Claim Pairs

• Qwen3.5-9B text-only achieves 95.5 ARC, 89.5 Align-FG, and 93.2 Overall, approaching GPT-5.1’s 95.9, 90.8, and 93.3.Structured claim fields do not consistently improve relation judgment and can introduce parsing noise, so Qwen3.5-9B is used downstream.Source: §5.1 · p. 5

5.2 Extracting Claims

• X-Claimify performs best across languages, especially on Exact-Aligned-F1, so it supplies reference claims for enrichment selection and downstream generation evaluation.Five adapted extraction frameworks are instantiated with GPT-5.1 and evaluated against three human reference claim sets per sentence using alignment-based F1 metrics.Source: §5.2 · pp. 5–6

5.3 Identifying Enrichment Claims

• Enrichment identification retrieves up to five candidates above cosine 0.7 before LLM verification, retaining additive, conflicting, and unmatched claims while discarding claims already fully covered.This permissive filter retains at least one candidate for 96% of target-side claims; conflicts remain candidates because Wikipedia sources may require contextual reconciliation.Source: §5.3 · p. 6
• Overall, 86.2% of non-English-side claims are selected as enrichment evidence, including 26.2% adding detail and 48.6% lacking relevant English counterparts.Only 13.9% are already covered by English.Source: §5.3 · p. 6

6 Enriching English Biographies

• The system preserves original English content while integrating non-English evidence through raw cross-lingual generation, machine translation, or structured enrichment claims.Claim-based enrichment filters already-covered information and uses Cadd as structured evidence for coherent rewriting rather than appending claims mechanically.Source: §6 · pp. 6–7, §6.1 · p. 7

6.2 Evaluation

• Generated biographies are evaluated against the original and target claim pools using claim growth, coverage, validity, hallucination, and Bal., which equally weights normalized supported growth and factual reliability.The evaluation also includes a human sanity check of readability, coherence, and Wikipedia-style writing on 45 biographies.Source: §6.2 · p. 8

6.3 Results and Analysis

• Claim-based evidence achieves the best average trade-off, raising #Valid to 20.47 while reducing hallucination to 23.54, compared with 19.05 and 29.59 for translation.Translation improves #Valid over raw biographies (15.89 versus 19.05) but leaves hallucination nearly unchanged (29.49 versus 29.59).Source: §6.3 · pp. 8–9
• French and Chinese outperform Azerbaijani, Mistral achieves the highest Bal. across language-evidence settings, and claim-based evidence obtains the highest readability and Wikipedia-style writing scores.Translation is the only method beating claims in Bal. for Azerbaijani with Mistral; claims score 3.933 for Readability/Fluency and 3.822 for Wikipedia-Style Writing.Source: §6.3 · p. 9

Conclusion

• The work introduces CLAW-4L, pairing English biographies of women with French, Chinese, and Azerbaijani biographies and resources for claim extraction and alignment evaluation.It also presents a framework that maps both biographies into a shared English claim space, selects non-English enrichment evidence, and performs controlled English rewriting.Source: §7 · p. 9
• Claim-based evidence offers the strongest trade-off between supported additions and hallucination in most language-generator settings.Compared with raw non-English biographies and translations, selected claims reduce long-context burden and improve interpretability.Source: §7 · p. 9
• Human evaluation finds no evident writing-quality penalty when selected claims provide rewriting evidence instead of other enrichment pipelines.The finding supports using selected claims as interpretable rewriting evidence.Source: §7 · p. 9
• Azerbaijani results show that lower-resource settings remain difficult, making robust claim extraction and alignment important for reliable cross-lingual enrichment.The conclusion identifies extraction and alignment robustness as key dependencies for lower-resource languages.Source: §7 · p. 9

Limitations

• CLAW-4L covers women’s biographies in French, Chinese, and Azerbaijani, enabling higher- and lower-resource comparisons but not representing Wikipedia’s full linguistic, regional, script, or biography diversity.The authors propose extending the benchmark to more languages and entity groups, including bidirectional and many-to-many enrichment beyond English.Source: Limitations · pp. 9–10
• Wikipedia claims may be erroneous, temporally mismatched, or incomplete, so the framework surfaces conflicts between language editions for downstream reconciliation rather than resolving factual truth.A proposed generate–verify–revise loop would remove or correct unsupported claims, retrieve independent evidence, expose provenance to editors, and abstain on unresolved conflicts.Source: Limitations · p. 10
• Automatic claim extraction and claim-pair judgment can propagate errors into enrichment, particularly in lower-resource languages, as reflected by missed Azerbaijani evidence.Human evaluation is also limited to 15 biography pairs, one generator, and three expert annotators, without direct comparison to original English biographies; GPT-5.1 and other closed-source steps may hinder exact reproduction.Source: Limitations · p. 10

B.1 Data Collection from Wikidata

• CLAW-4L is built from Wikidata-eligible female entities with English and corresponding French, Chinese, or Azerbaijani sitelinks, then cleaned to retain readable biography prose while removing markup and non-biographical sections.The candidate pools contain 3,434 entities overall, including 1,500 each for France and China and all 434 eligible Azerbaijani entities.Source: §B.1 · p. 13, §B.2 · p. 13

B.3 Length-based Tiering and Language Calibration

• Language-calibrated normalized token ratios provide a cross-lingual content-richness heuristic, with inflation factors above 1 for French and Azerbaijani and below 1 for Chinese; candidates are then tiered rather than hard-filtered.The benchmark prioritizes Tiers A and B, uses Tier C as a fallback, and applies approximate occupation-aware sampling across artist, athlete, scientist, and politician groups.Source: §B.3 · p. 14, §B.5 · p. 14

B.6 Stratified Biography Sampling for CLAW-4L

• The final 300-instance CLAW-4L closely matches the 900-instance source pool, exactly preserving occupation and selection-tier distributions and nearly matching remaining length, nationality, and infobox distributions.Constrained local search enforces occupation counts and optimizes other attributes jointly; extracted GPT-5.1 claims then support sentence selection and structured human annotation for CLAW-4L-CX.Source: §B.6 · pp. 15–16, §B.7 · p. 16

B.8 Sentence Selection for CLAW-4L-CX

• Candidates are partitioned into 1-, 2-, 3-, and 4+-claim buckets, with equal 0.25 proportions enforced globally to preserve complexity diversity.This prevents independent top-scoring selection from over-favoring simple one-claim sentences and retains medium- and high-complexity examples.Source: §B.8 · pp. 18–19
• Sentence quality combines claim coverage, fact density, and rewrite cost through equal-weighted within-biography rank aggregation.Coverage measures faithfulness to the source sentence, density captures structured informational richness, and rewrite cost reflects contextual dependence or reformulation difficulty.Source: §B.8 · pp. 18–19
• A constrained optimization maximizes total sentence quality while selecting exactly one sentence per biography and meeting target claim-count bucket totals.The selected set remains close to the sampled pool’s natural complexity distribution while avoiding trivial-sentence over-representation.Source: §B.8 · pp. 19–20
• The procedure yields 300 English and 300 non-English sentences for CLAW-4L-CX, selected independently so paired biographies are not sentence-aligned.The resulting benchmark is a sentence-level human annotation resource.Source: §B.8 · p. 20

B.9 CLAW-4L-CX Statistics

• CLAW-4L-CX contains 300 English and 300 non-English selected sentences spanning claim-count buckets, with means of 2.63 and 2.72 claims, respectively.Each language side has 25 instances per country in the 1-, 2-, 3-, and 4+-claim buckets, preserving both simple and claim-dense cases.Source: §B.9 · p. 20, Table 11 · p. 23

B.10 CLAW-4L-CX Annotators and Interface

• Annotations were produced by uncompensated volunteer researchers, while post-hoc agreement compares annotator claim sets using classifier-scored aligned and exact-aligned cross-set pairs.Agreement is summarized with average precision, recall, and F1 across annotator pairs for multiply annotated sentences in each language.Source: §B.10 · p. 20, §B.11 · p. 20

B.12 Construction of CLAW-4L-RC

• CLAW-4L-RC is a 600-instance English claim-pair benchmark built from normalized cross-lingual claim extractions, with structured fields and manually validated relation labels.LaBSE similarity narrows candidate alignments, diversity filtering balances countries and claim patterns, and validation checks both textual and structured content.Source: §B.12 · p. 22
• Each claim pair is labeled Aligned when both claims describe the same factual aspect and can both be true, Contradicted when they assert incompatible values or states, or Not Relevant when they describe different factual aspects.Only Aligned pairs receive an enrichment label; A=B means neither claim contains meaningful extra atomic factual content beyond the other.Source: §B.12 · p. 22

B.9 CLAW-4L-CX Statistics

• CLAW-4L-CX selection improves normalized coverage, fact-density, rewrite, and overall ranks for both English and non-English sentences, enriching for higher-quality claim-bearing examples.English shifts are +0.4107, +0.2095, +0.1965, and +0.2722; non-English shifts are +0.3285, +0.2221, +0.4030, and +0.3179, respectively.Source: §B.9 · p. 20, Table 12 · p. 23

B.13 CLAW-4L-RC Statistics

• CLAW-4L-RC contains 400 aligned, 100 contradicted, and 100 not-relevant pairs, with each of the four aligned sub-labels balanced at 100 instances.Across BGE and MPNet embeddings, exact alignment is most similar, not-relevance is substantially lower, and contradiction remains relatively close because it can share topical or lexical content.Source: Table 14 · p. 24, §B.13 · p. 24

C.1 Models

• Classifier evaluation compares instruction-tuned open-source models, GPT-5.1, and NLI baselines using text-only or text-plus-infobox inputs and accuracy aggregates for fine-grained, top-level, and overall relations.The six labels are evaluated with Align-FG, ARC, and Overall, which respectively emphasize aligned distinctions, top-level relation types, and equal weighting across all labels.Source: §C.1 · pp. 24–25, §C.2 · p. 25

C.3 NLI Baseline Adaptation

• NLI baselines are adapted bidirectionally into the CLAW-4L-RC label space but cannot fully represent all six relations, so only directly computable label-wise accuracies are reported.MiniCheck-7B collapses mutual partial alignment, contradiction, and non-relevance into unsupported outcomes, while DeBERTa-v3-large-mnli cannot distinguish mutual partial alignment from non-relevance and omits aggregate scores.Source: §C.3 · p. 25

Overall classifier performance

• GPT-5.1 achieves the strongest aggregate performance, while Qwen3.5-9B is the closest open-source model across ARC, Align-FG, and Overall.Qwen3.5-9B scores 95.5 versus GPT-5.1’s 95.9 on ARC and substantially outperforms other open-source families.Source: Figure 9 · p. 26, §C.4 · p. 26

NLI baselines

• NLI baselines transfer well to contradiction and exact alignment but cannot represent finer enrichment relations, especially mutual partial alignment.DeBERTa-v3-large-mnli reaches 95.0 on contradiction and 90.0 on exact alignment, yet label mismatch limits its usefulness as a full classifier.Source: §C.4 · pp. 25–26

Fine-grained alignment

• Align-FG scores are consistently lower than ARC, with mutual enrichment A↔B typically the hardest distinction despite a positive correlation between the metrics.The challenging distinctions include A>B, B>A, and A↔B, requiring precise reasoning about bidirectional non-conflicting information.Source: §C.4 · p. 26

Structured infoboxes

• Infobox augmentation often slightly harms stronger models but occasionally helps smaller ones, indicating that its value depends on capacity and input fidelity.For stronger models including Qwen3-14B, Qwen3.5-9B, and GPT-5.1, structured inputs can add noise or redundancy.Source: §C.4 · p. 26

Scaling trends

• Larger models generally perform better within Qwen3 and Qwen3.5, whereas Ministral-3 reverses this trend; these results motivate Qwen3.5-9B as the default open-source classifier.The Ministral-3 3B model outperforms its larger variants.Source: §C.4 · p. 26

D.1 Core Operations in Claim Extraction

• Claim extraction combines decomposition, decontextualization, verifiable selection, and disambiguation to produce independently checkable factual units.These operations split compound statements, make claims self-contained, filter non-verifiable material, and resolve or avoid ambiguous references.Source: §D.1 · p. 28

D.2 Adapted English Claim Extraction Frameworks

• Five English extraction methods are adapted to accept English or non-English sentences and output atomic factual claims in English for shared-space alignment.The adapted methods are X-FactScore, X-FactCheck-GPT, X-DnDScore, X-VeriScore, and X-Claimify.Source: §D.2 · p. 28

D.3 Evaluation on CLAW-4L-CX

• CLAW-4L-CX evaluation uses classifier-based claim-set matching against three human reference sets, with degree-normalized precision, recall, and F1 under aligned and exact criteria.Fan-out normalization penalizes overly broad or redundant claims, while Aligned-F1 measures semantic recovery and Exact-Aligned-F1 emphasizes specificity.Source: §D.3 · p. 29

D.4 Two-stage Selection Protocol

• The two-stage protocol first compares extraction frameworks with GPT-5.1, then compares open-source backbones after fixing the best framework.The second stage evaluates Qwen3.6-27B, Gemma-4-31B-it, and Mistral-3.2-24B-it for cost, reproducibility, and scalability.Source: §D.4 · pp. 29–30

D.5 Final Choice

• X-Claimify achieves the strongest Exact-Aligned-F1 across English, French, and Chinese, while also attaining the best or near-best Aligned-F1.These results motivate its use for downstream alignment because it preserves source content and comparable event-level granularity.Source: §D.5 · p. 30, Table 17 · p. 31
• 82.62 versus 78.05 on French Aligned-F1, but 70.18 versus 73.56 on Exact-Aligned-F1, leaves Gemma-4-31B-it weaker than GPT-5.1 for exact claim formulation.GPT-5.1 remains the strongest and most consistent backbone overall, so GPT-5.1 + X-Claimify is retained for all languages.Source: §D.5 · p. 30, Table 18 · p. 31

E Machine Translation Model Selection for Biography Enrichment

• Translation selection compares three open-source generator LLMs with M2M100-1.2B and NLLB-200-3.3B on nine long biographies spanning French, Chinese, and Azerbaijani.The subset contains the three biographies with the largest token counts in each language, targeting context-length and information-density challenges.Source: E Machine Translation Model Selection… · p. 30

E.1 Translation Protocol

• Translations are generated incrementally with a 4,096-token non-English input cap, section-based chunking, newline and token-level fallback splitting, and final chunk concatenation.Translation uses only the non-English biography as input, allowing a larger per-chunk budget than the generation setting.Source: §E.1 · pp. 30, 32

• The protocol extracts reference and translated claims with X-Claimify plus GPT-5.1, retrieves candidates with MPNet, and verifies filtered pairs using degree-normalized scoring.Retrieval keeps top-five neighbors above cosine similarity 0.7, while Aligned-F1 tolerates granularity differences and Exact-Aligned-F1 requires atomic claim matching.Source: §E.2 · p. 32
• Exact-Aligned-F1 is consistently much lower than Aligned-F1 across language pairs and models, making exact claim preservation substantially more demanding.The strict metric is sensitive to paraphrasing, granularity shifts, and small factual perturbations, although many claims remain semantically relevant.Source: §E.2 · p. 32
• The three open-source LLMs consistently outperform M2M100-1.2B and NLLB-200-3.3B across all language pairs under both alignment metrics.The passage attributes this pattern to larger model scale and better handling of long, discourse-dependent biography context.Source: §E.2 · p. 32
• French-to-English is easiest and Chinese-to-English hardest; final selections are Gemma-4-31B-it for Azerbaijani, Mistral-3.2-24B-it for French, and Qwen3.6-27B for Chinese.French has coarse scores of 93.80 versus 93.77 but exact scores of 48.98 versus 45.48; Chinese selection prioritizes semantic preservation over exact-form matching.Source: §E.2 · p. 32

F Target-Side Enrichment Claim Selection

• The selector retrieves up to five English candidates per non-English claim when cosine similarity reaches 0.7, avoiding exhaustive claim-pair verification.Top-5 is chosen as a middle operating point between broader candidate coverage and verifier-call cost.Source: F Target-Side Enrichment Claim Selection · pp. 32–33
• Non-English-more-specific, mutually complementary, conflicting, and unmatched claims are retained as enrichment candidates, while exact and English-more-specific matches are excluded.Conflicts require contextual reconciliation, and the generator must integrate only compatible, supported additions with the original English biography.Source: F Target-Side Enrichment Claim Selection · p. 33
• 89.3 non-English claims per biography pair (86.2%) are selected as enrichment candidates, while only 13.9% are already covered by English claims.Target-additive claims account for 26.2%, unmatched claims 48.6%, and conflicts 11.4%.Source: F Target-Side Enrichment Claim Selection · pp. 33–34, Table 21 · p. 34

G.1 Evaluation Metrics

• Generated claims are evaluated against the combined English and non-English reference pool, with contradiction taking priority over validity and halluc.% combining unsupported and contradicted claims.The evaluation also measures claim growth, estimated supported gain, and the overall Bal. metric.Source: §G.1 · p. 34

G.2 Full Results

• The main analysis averages supported additions and hallucination rates across three non-English languages and three generators for each evidence format.Table 28 provides the full results across language, evidence-format, and generator combinations.Source: §G.2 · p. 34, Table 22 · p. 34

G.3 Qualitative Error Analysis

• Qualitative analysis inspects 18 biographies spanning three languages, two contrasting support cases per language, and the Raw, Translation, and Claims methods.Cases are selected using verifier-supported output claims under the best-performing format for each language, while error types come from manual inspection.Source: §G.3 · p. 34, Table 23 · p. 35
• The qualitative examples illustrate recurring omission and hallucination patterns, but their high- and low-support sampling does not establish error prevalence in the full benchmark.Table 24 presents representative cases using evidence from the reference pool and generated biographies or their claim decompositions.Source: §G.3 · p. 35, Table 24 · p. 35

G.4 Human Evaluation of Writing Quality

• Three English-working Ph.D. researchers independently rate 45 outputs on readability, coherence, and Wikipedia-style writing using 1–5 Likert scales, with method labels visible during annotation.The evaluation is treated as a writing-quality sanity check rather than a definitive editorial assessment because it was not method-blind.Source: §G.4 · p. 35
• Claims achieves the highest scores on all three writing-quality dimensions in the longest bucket, although this pattern is descriptive because each method–bucket cell contains only three outputs.The longest bucket includes the 6,384-word repetitive Azerbaijani Translation output, which qualifies interpretation of the result.Source: §G.4 · p. 35, Table 26 · p. 36
• Ordinal Krippendorff’s α ranges from 0.507 to 0.537, while three-rater mean reliability reaches ICC(2,3) values of 0.801–0.817.ICC(2,1) measures single-rating reliability, whereas ICC(2,3) measures reliability of the mean of three ratings.Source: §G.4 · pp. 35–36, Table 27 · p. 36

H Main Prompt

• The appendix reports the main prompts because space is limited and provides the complete prompt set, including variants for adapted extraction frameworks and enrichment-generation methods, in a public repository.Source: H Main Prompt · p. 36

H.1 Prompt for X-Claimify

• X-Claimify is prompted through separate selection, disambiguation, and decomposition stages, with the corresponding instructions shown in Figures 10–19.Source: §H.1 · p. 36, Figure 10 · p. 38, Figure 11 · p. 39

H.2 Prompt for LLM classifier

• The LLM classifier uses a system prompt and a separate user prompt, provided in Figures 22–24.Source: §H.2 · p. 36, Figure 22 · p. 46, Figure 23 · p. 47

H.3 Prompt for "English + non-English" Generation

• The cross-lingual generation materials provide a system prompt and a user prompt for improving an existing English biography with additional facts, presented in Figures 25 and 26.Source: §H.3 · p. 36, Figure 25 · p. 48, Figure 26 · p. 49

H.4 Prompt for LLM-based Biography Translation

• The biography-translation prompt instructs the model to produce natural, faithful English while preserving all facts, dates, names, numerals, and structure without additions or omissions, using the same prompts for all three open-source LLMs.Source: §H.4 · p. 36, Figure 27 · p. 50