PaperlayerView exact revision on arXiv

Source-linked AI summary

Figurative and Cultural Knowledge in LLMs: Investigating Cross-Domain Transfer through Fine-Tuning

Mena Attia, Mona Diab, Thamar Solorio

arXiv:2608.18361v1cs.CL

TL;DR

Cultural knowledge and figurative understanding are often studied separately, leaving their relationship in LLMs uncertain. This paper tests bidirectional transfer through controlled Arabic fine-tuning and finds transfer generally limited and model-dependent, with poetry fine-tuning improving idiom interpretation (+2.33%, p = 0.021).

  • Problem

    Whether cultural knowledge and figurative understanding support one another in large language models remains untested, despite being evaluated separately.

    Source: §1 · p. 2

  • Method

    The study uses controlled fine-tuning and evaluation experiments in Arabic to test bidirectional transfer between cultural knowledge and figurative understanding.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Cross-domain transfer was limited, inconsistent, and model-dependent; poetry fine-tuning on FannOrFlop improved idiom interpretation (+2.33%, p = 0.021).

    Source: §6 · p. 8

  • Takeaways & Limitations

    The relationship between cultural and figurative knowledge is not straightforwardly captured through fine-tuning alone.

    Source: §1 · p. 2

    The transfer directions use different source datasets, target benchmarks, and task formulations, so their magnitudes cannot establish directional asymmetry.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Figurative language is deeply culturally embedded; fluent use requires not just linguistic competence but cultural immersion. We ask whether LLMs can learn this link: does fine-tuning on cultural data improve figurative language understanding, and vice versa? We conduct a systematic study across four models (ALLaM-7B, Fanar-1-9B, Qwen3-8B, Llama-3.1-8B) and six Arabic datasets spanning cultural commonsense, proverbs, and poetry across diverse dialects and regions. Fine-tuning on poetry improves idiom comprehension (+2.33%, p<0.05), a gain our ArabicMMLU control does not reproduce, indicating that it stems from figurative content rather than Arabic language adaptation and pointing to a sensitivity to non-literal meaning that transfers across figurative types. Cultural fine-tuning, by contrast, lowers proverb-interpretation accuracy in both Arabic-centric models. Transfer between the two domains is otherwise indistinguishable from noise, with Arabic models frequently regressing after fine-tuning, suggesting prior saturation of relevant knowledge, while multilingual models show greater adaptation headroom. Error analysis further reveals that fine-tuning reinforces experiential cultural knowledge while destabilizing historically grounded factual knowledge. Our findings suggest that the relationship between culture and figurative language, though conceptually natural, is not straightforwardly captured through fine-tuning alone.

1 Introduction

The introduction frames figurative competence and cultural knowledge as coupled, then investigates their transfer through controlled Arabic fine-tuning; reported transfer is limited and model-dependent, with poetry the only reliable positive source.

1 Introduction

  • Figurative meanings encode community-specific values and experience, so interpretation depends on cultural familiarity as well as linguistic competence (Kövecses, 2005; Sharifian, 2015; Dobrovol’skij and Piirainen, 2022; Mieder, 2004, 1993; Gibbs, 1994).Source: §1 · pp. 1–2
  • NLP benchmarks study cultural knowledge and figurative interpretation separately, leaving whether either supports the other in large language models untested.Source: §1 · p. 2
  • The study tests cultural-to-figurative and figurative-to-cultural transfer, cross-form transfer, and whether gains reflect content or Arabic-language adaptation.Controlled Arabic fine-tuning examines cultural data against idiom and proverb interpretation, figurative data against culturally grounded question answering, and transfer to unseen figurative forms.Source: §1 · p. 2
  • Transfer effects are generally small, inconsistent, and model-dependent; poetry fine-tuning is the only reliable source of positive figurative transfer.The authors describe this as evidence that cross-form figurative transfer may extend beyond a single non-literal category.Source: §1 · p. 2

2 Related Work

Related work separates figurative-language benchmarks from largely factual cultural evaluation, while prior adaptation studies address figurative interpretation or cultural knowledge rather than transfer between them.

2.1 Benchmarks for Figurative Language and Culture

2.2 Improving Figurative Interpretation

  • Work on figurative interpretation chiefly uses task-specific fine-tuning or structured prompting, with smaller efforts adding contextual inference and compositional reasoning.Prior studies include cohesion and translation objectives for idiom detection, prompting for metaphor and pragmatic reasoning, and contextual inference strategies for idioms and similes.Source: §2.2 · p. 3

2.3 Cultural Adaptation and Cross-Domain Transfer

  • Cultural adaptation targets cultural knowledge through fine-tuning or culture-specific adapters; this study instead tests transfer between culture and figurative tasks and from poetry to proverbs and idioms.For Arabic, the PalmX shared task (Alwajih et al., 2025b) has driven cultural fine-tuning systems, and Almheiri et al. (2025) report cross-country transfer under lightweight alignment.Source: §2.3 · p. 3

3 Experimental Setup

The study uses controlled fine-tuning experiments to test transfer between cultural and figurative-language understanding in Arabic LLMs, with a general-Arabic control to separate content effects from language exposure. It compares multilingual and Arabic-focused models using shared datasets and evaluation conditions.

3 Experimental Setup

  • Three experiments test cultural fine-tuning, figurative fine-tuning, and general-Arabic fine-tuning for transfer to idiom and proverb understanding.Cultural data is evaluated on idioms and proverbs; poetry and proverbs are evaluated on the same benchmarks, while general Arabic data tests whether transfer reflects content or Arabic exposure.Source: §3 · p. 3

3.1 Datasets

  • Training uses cultural and figurative Arabic resources, while evaluation uses AraDiCE, Jawaher, and Kinayat; ArabicMMLU supplies the language-content control.The datasets cover Arabic varieties and regions; ArabicMMLU control training uses its Arabic Language and Grammar subsets (980 examples total).Source: §3.1 · pp. 3–4, Table 1 · p. 4
  • Cultural training excludes Palm proverb-topic items, and each condition uses 1,000 samples except Jawaher, which has 800 training samples.The exclusions prevent cultural conditions from receiving figurative supervision; full-corpus FannOrFlop and ArabCulture training is reserved for an appendix ablation.Source: §3.1 · p. 3
  • Training and evaluation are framed as open-ended generation, with correct answers or explanations as targets; Jawaher uses Attia et al. (2026) distractors.Jawaher and Kinayat have two options per item (50% chance baseline), while AraDiCE has three (33.3%).Source: §3.1 · p. 4

3.2 Models and Evaluation

  • Four instruction-tuned models—two multilingual and two Arabic-centric—are fine-tuned with LoRA and evaluated zero-shot under identical conditions.Evaluations use three random-seed runs, with answer-option shuffling; accuracy is averaged across runs, scoring multiple-choice answers by option log-likelihood.Source: §3.2 · p. 4

4 Results

Poetry fine-tuning produced the only reliable aggregate gain, on idiom interpretation; other cross-domain effects were generally small, noisy, and model-dependent.

4 Results

  • 2.33% was the average Kinayat gain from FannOrFlop fine-tuning, the study’s only pooled effect with a 95% confidence interval excluding zero.The gain reflects transfer between figurative-language tasks, not culture-to-figurative transfer.Source: §4 · p. 4, §4.2 · p. 6
  • −3.70% and −3.03% were the Jawaher declines for ALLaM-7B and Fanar-1-9B, respectively, under Palm fine-tuning—the only statistically supported negative effects.Both declines had the same sign across all three seeds.Source: §4 · pp. 4–5

4.1 Transfer between Cultural and Figurative Language Fine-tuning

  • Cultural fine-tuning did not reliably improve figurative benchmarks, while Arabic-centric and multilingual models showed opposing responses to Palm fine-tuning.ALLaM-7B and Fanar-1-9B regressed across all three evaluation datasets; LLaMA-3.1-8B had positive point estimates, while Qwen3-8B had one significant Kinayat seed gain (+4.67%, p = 0.016).Source: §4.1 · p. 5
  • Proverb and poetry fine-tuning produced no detectable pooled AraDiCE improvements, and the overall −1.85% ALLaM-7B FannOrFlop decline was driven by one significant seed.That seed declined 5.00% (p = 0.022), while the other two seeds changed by +1.11% and −1.67%.Source: §4.1 · pp. 5–6
  • ArabicMMLU fine-tuning lowered Kinayat and Jawaher performance on average but improved AraDiCE for all four models.The Kinayat control result moved opposite to poetry fine-tuning’s reliable gain; because ArabicMMLU is natively produced Arabic, the contrast is curated versus incidental cultural exposure, not cultural versus non-cultural data.Source: §4.1 · p. 6

4.2 General Observations

  • Kinayat was the most fine-tuning-responsive benchmark, while Arabic-centric models more often regressed and multilingual models more often had positive, usually nonsignificant estimates.Across fine-tuning datasets, average effects ranged from −0.46% to +2.33%; most were statistically indistinguishable from zero, consistent with limited cross-task transfer.Source: §4.2 · p. 6

5 Error Analysis

Error analysis finds near-zero aggregate cultural-QA change despite item-level and cultural-context redistribution, while poetry fine-tuning improves idiom interpretation. Where predictions shift, they favor experiential cultural knowledge over historically grounded or domain-specific facts.

5.1 Figurative Language and Cultural Understanding

  • Across models and seeds, cultural QA gains (185) and regressions (189) yield net −4 and average change −0.11%, consistent with a null pooled effect.These aggregate patterns are descriptive, not individually tested; counts are pooled across models and seeds, and some categories contain few items.Source: §5.1 · pp. 6–7
  • Fanar and Qwen have small positive cultural-QA point estimates under both fine-tuning sets, while ALLaM degrades most under poetry (−1.87%); LLaMA is mixed.ALLaM’s poetry estimate rests on one significant seed (−5.00%, p = 0.022) against +1.11% and −1.67% in the other seeds, so its direction is better supported than its magnitude.Source: §5.1 · p. 7, Table 3 · p. 7
  • Jordan shows the strongest country net gain (+32), while Syria (−18) and Egypt (−12) decline; topic gains favor Food/Cuisine (+15) and Traditional Games (+13), unlike clothing, history, and religion.Traditional Clothing records −14, while History/Civilization and Religion each record −8; the latter two are based on few items. Overall, shifts favor experiential cultural knowledge over historical or domain-specific facts.Source: §5.1 · p. 7, Table 4 · p. 7

5.2 Culture and Poetry as Signals for Figurative Language

  • Cultural fine-tuning shifts idiom accuracy by +1.00% and proverb accuracy by −0.27%, neither distinguishable from zero; poetry effects are smaller on proverbs.Proverb effects under both conditions may be smaller because the proverb set has greater dialectal diversity.Source: §5.2 · p. 8, Table 5 · p. 8
  • Poetry fine-tuning improves idiom accuracy by +2.33%; its +1.01% proverb effect is not significant, supporting transfer of non-literal sensitivity across figurative types.The evidence for broader figurative competence rests on the poetry-to-idiom result; the proverb effect does not reach significance (p = 0.23).Source: §5.2 · pp. 7–8, Table 5 · p. 8
  • Under cultural fine-tuning, ALLaM and Fanar regress while LLaMA and Qwen improve; poetry also benefits Fanar on both tasks, unlike cultural fine-tuning.The stronger Arabic-centric models’ greater proverb harm is interpreted as possible noise from fine-tuning on a narrower curated corpus relative to knowledge acquired during pretraining.Source: §5.2 · p. 8

6 Conclusion

Across four models and six Arabic datasets, cross-domain transfer was limited, inconsistent, and model-dependent. Poetry fine-tuning produced the clearest reliable gain, while cultural fine-tuning could regress proverb interpretation in Arabic-centric models.

6 Conclusion

  • Cross-domain transfer was limited: cultural tuning yielded small, unreliable Kinayat gains (+0.78% ArabCulture; +1.22% Palm), while figurative tuning showed no detectable cultural-benchmark change.The pooled cultural-tuning gains masked polarized per-model effects.Source: Table 6 · p. 8, §6 · p. 8
  • Palm fine-tuning produced the study’s only statistically supported regressions, lowering both Arabic-centric models’ accuracy on Jawaher.Cultural-training effects on figurative benchmarks were polarized in sign across models.Source: §6 · p. 8
  • Poetry fine-tuning improved idiom interpretation by +2.33% (p = 0.021), while the ArabicMMLU control regressed on that benchmark, implicating figurative content rather than added language exposure.This was the study’s one reliable effect within the figurative domain, not across the cultural–figurative divide.Source: §6 · p. 8
  • Arabic-centric models started from higher baselines and accounted for both supported regressions, whereas multilingual models showed greater headroom for adaptation.The authors suggest Arabic-centric models may already have internalized much of the relevant knowledge during pre-training.Source: §6 · p. 8
  • Fine-tuning favored culturally embedded experiential knowledge over historically grounded or domain-specific facts; small evaluation sets make these findings a constraint on current Arabic resources, not proof no relationship exists.Examples of experiential knowledge include food, customs, and celebrations.Source: §6 · p. 8

Limitations · A Prompt Templates · B Fine-tuning Setup

The appendices describe the study’s methodological limitations, task-specific prompt templates, and LoRA fine-tuning configuration. The limitations concern adaptation and hyperparameter scope, model scale, and the comparability and generality of cross-domain transfer claims.

Limitations

  • Experiments used LoRA and limited hyperparameter search on 7–9B models, so alternative adaptation methods, better tuning, or larger models could produce different results.The authors leave broader exploration of model sizes, fine-tuning strategies, and hyperparameter configurations to future work.Source: Limitations · p. 9
  • Transfer directions are not directly comparable because they use different datasets, benchmarks, and task formulations, and neither domain reliably generalizes to the other across models and conditions.Differences in magnitude may reflect dataset difficulty, domain specificity, or evaluation sensitivity, so the authors avoid conclusions about inherent directional strength.Source: Limitations · p. 9

A Prompt Templates

  • MCQ understanding uses a question with candidate answers and a single-letter response, while AraDiCE uses the same structure with three options.Figure 3 depicts the MCQ prompt and Figure 4 the AraDiCE-Culture template.Source: A Prompt Templates · p. 12, Figure 3 · p. 12, Figure 4 · p. 12
  • FannOrFlop, Jawaher, and ArabCulture fine-tuning prompts pair a task-specific preamble with a poem, proverb, or cultural prompt and request a free-form response.The proverb prompt asks for the correct explanation and instructs the model to output only its letter; Figures 5–7 show the respective templates.Source: A Prompt Templates · p. 12, Figure 5 · p. 12, Figure 6 · p. 12
  • The Palm instruction-following prompt omits a task-specific preamble, relying on the question field alone to provide the context for response generation.Figure 8 shows the Palm template.Source: A Prompt Templates · p. 12, Figure 8 · p. 12

B Fine-tuning Setup

  • LoRA was applied to attention and feed-forward projection layers at rank 4, with α = 8 and dropout 0.1; training ran for 3 epochs at 5e −5 with batch size 1.A validation set monitored convergence; base parameters remained frozen, and fine-tuning used one NVIDIA RTX 5000 Ada Generation GPU.Source: B Fine-tuning Setup · pp. 12–13

C Additional Results · D Ablations

Additional results report three-seed accuracy evaluations and visualize how fine-tuning on four datasets and an ArabicMMLU baseline changes accuracy relative to base models.

C Additional Results

  • Table 7 reports accuracy across three random-seed runs on Jawaher, Kinayat, and AraDiCE for base and subset-fine-tuned models.The multiple runs assess whether observed trends are stable under random initialization.Source: C Additional Results · p. 13
  • Figure 9 visualizes held-out dataset accuracy differences after fine-tuning on ArabCulture.The difference is fine-tuned accuracy minus base accuracy.Source: Figure 9 · p. 13, C Additional Results · p. 13
  • Figure 10 visualizes held-out dataset accuracy differences after fine-tuning on FannOrFlop.The difference is fine-tuned accuracy minus base accuracy.Source: Figure 10 · p. 13, C Additional Results · p. 13
  • Figure 11 visualizes held-out dataset accuracy differences after fine-tuning on Jawaher.The difference is fine-tuned accuracy minus base accuracy.Source: Figure 11 · p. 13, C Additional Results · p. 13
  • Figure 12 visualizes held-out dataset accuracy differences after fine-tuning on Palm.The difference is fine-tuned accuracy minus base accuracy.Source: Figure 12 · p. 13, C Additional Results · p. 13
  • Figure 13 visualizes dataset accuracy differences after fine-tuning on the ArabicMMLU baseline.The difference is fine-tuned accuracy minus base accuracy.Source: Figure 13 · p. 13

D.1 Increasing the size of the fine-tuning data · D.2 Changing LoRA hyperparameters

Increasing fine-tuning data produced model- and domain-dependent results, with ALLaM the most consistent beneficiary; alternative LoRA settings yielded only partial gains and remained below the base model. The findings are reported for the two appendix sections on data size and LoRA hyperparameters.

D.1 Increasing the size of the fine-tuning data

  • The data-size comparison tested full FannOrFlop and ArabCulture datasets against the default 1,000-sample subsets.The full datasets contained 6,984 and 3,482 samples, respectively.Source: §D.1 · p. 13, Table 8 · p. 15
  • Full-dataset fine-tuning consistently benefited ALLaM across FannOrFlop and ArabCulture, but other models did not improve uniformly.The comparison uses full datasets of 6,984 FannOrFlop and 3,482 ArabCulture samples versus the default 1,000-sample subset.Source: §D.1 · pp. 13–14, Figure 14 · p. 14, Figure 15 · p. 14
  • On FannOrFlop, ALLaM and Qwen improved across all three benchmarks, while Fanar fell most on Jawaher (−3.37%) and LLaMA on Kinayat (−8.44%).The benchmarks were Jawaher, Kinayat, and AraDiCE.Source: §D.1 · pp. 13–14, Figure 14 · p. 14, Table 8 · p. 15
  • On ArabCulture, ALLaM improved modestly across all benchmarks and Fanar gained 1.48% on AraDiCE, while LLaMA fell 3.20% on Jawaher and 4.22% on Kinayat.Qwen also declined, though no specific value is reported in the passage.Source: §D.1 · p. 14, Figure 15 · p. 14, Table 8 · p. 15

D.2 Changing LoRA hyperparameters

  • With ALLaM fine-tuned on Palm Subset, LoRA rank r = 16, α = 32, and lr=1e-5 improved Jawaher by 1.18% and AraDiCE by 2.04% over the default configuration.The passage describes these as the only notable gains among the alternative settings.Source: §D.2 · p. 14, Table 9 · p. 15
  • The r = 16, α = 32, lr=5e-4 setting gained 2.59% on AraDiCE alone, and all alternative LoRA settings remained below base ALLaM performance.Overall, the alternative hyperparameter settings largely failed to improve on the default fine-tuned configuration.Source: §D.2 · pp. 14–15

E Confidence Intervals and Significance Testing. · F Error Analysis · F.1 Using Figurative Language to Improve Cultural Understanding

Across pooled tests, poetry fine-tuning significantly improved Kinayat idiom accuracy, while other pooled effects were indistinguishable from noise. Error analysis frames transfer from figurative-language fine-tuning as selective, with cultural questions varying in receptivity and degradation risk.

E Confidence Intervals and Significance Testing.

  • Paired 95% confidence intervals use 20,000 item-index bootstrap resamples, with exact McNemar tests authoritative when fewer than ten predictions differ.The paired resampling preserves the correspondence between base and fine-tuned predictions.Source: E Confidence Intervals and… · p. 15
  • The cluster bootstrap resamples items jointly across model-by-seed measurements, so it does not require repeated measurements to be independent.The reported intraclass correlations were small (all |ICC| < 0.08), making the clustering correction modest in its effect on interval width.Source: E Confidence Intervals and… · p. 15
  • Across most conditions, fewer than 20 predictions changed in either direction and intervals were wide and usually crossed zero.Several Fanar and Qwen seeds had fewer than ten changes, the sparse case in which the exact test is authoritative.Source: E Confidence Intervals and… · p. 15
  • Only five of 72 per-seed tests reached p < 0.05, close to the global-null expectation, so substantive claims rely on pooled estimates rather than individual runs.The authors treat per-cell results as descriptive.Source: E Confidence Intervals and… · p. 15

F Error Analysis

  • Table 13 lists the ten most frequently improved and regressed questions from the AraDiCE cultural benchmark.The table organizes questions by how often they improved or regressed across models fine-tuned on figurative-language data.Source: §F.1 · pp. 15–16

E Confidence Intervals and Significance Testing.

  • +2.33 percentage points (p = 0.021) was the only significant pooled effect: poetry fine-tuning improved Kinayat idiom accuracy; all other pooled effects were within noise of zero.AraDiCE pooled estimates were essentially null, and all |ICC| values were below 0.08.Source: E Confidence Intervals and… · p. 15, Table 11 · p. 16

F.1 Using Figurative Language to Improve Cultural Understanding

  • The error analysis examines which cultural-knowledge categories are receptive to transfer from figurative-language fine-tuning and which are prone to degradation.It presents this as a granular view of cross-domain transfer's asymmetric and selective nature.Source: §F.1 · p. 16

F.2 Using Culture and Poetry to Improve Figurative Language Understanding

Fine-tuning transfer varies by dialect and individual expression: cultural and poetry fine-tuning produce opposing proverb outcomes across varieties, while item-level results are selective and unstable. Persistent dialect patterns suggest evaluation-set properties may matter more than fine-tuning data choice.

Dialect-Level Analysis

  • Fine-tuning transfer is selective and unstable at the item level: 21 idioms and 15 proverbs improve in some seeds but regress in others.The listed improved and regressed expressions identify items with consistent gains or degradation, while some expressions behave inconsistently across fine-tuning conditions and models.Source: §F.2 · p. 16, Table 14 · p. 17, Table 15 · p. 17
  • Cultural fine-tuning yields the sharpest proverb regressions for Algeria (−11), Sudan (−10), and Libya (−8), but gains for Mauritania (+20), Yemen (+14), Iraq (+5), and Tunisia (+5).The authors note that Algeria, Sudan, and Libya were represented in both cultural fine-tuning datasets, while Mauritania and Iraq were not represented in ArabCulture, leaving the geographic split unexplained.Source: §F.2 · p. 16
  • Under poetry fine-tuning, Sudanese (−8), Qatari (−8), and Algerian (−5) proverbs regress most, while Yemeni (+8), Mauritanian (+8), and Jordanian (+7) benefit most.Qatari proverbs shift from mildly negative under cultural fine-tuning to a clear poetry-fine-tuning regression, largely driven by one proverb; Egyptian proverbs shift from −5 to +3, possibly because poetry reinforces Egyptian Arabic figurative patterns more directly.Source: §F.2 · pp. 16–17
  • Persistent Mauritanian and Yemeni gains alongside Algerian and Sudanese losses suggest dialect difficulty reflects evaluation-set properties more than fine-tuning-data choice.The authors suggest the underlying difficulty is linked to intrinsic properties of those dialects in the evaluation set, rather than the choice of fine-tuning data.Source: §F.2 · p. 17

G Use of AI Assistants · ¯ ÈAªË@ · àA¿ ñË ð Q J.ºË@ ÐC¿ ©ÖÞ @

The section discloses that AI assistants supported manuscript writing and editing, while the authors retained responsibility for the research and final wording. It also catalogs item-level fine-tuning improvements and regressions across idioms, proverbs, and dialects.

G Use of AI Assistants

  • AI assistants supported manuscript rephrasing and editing, but authors made all research-design, analysis, interpretation, and final-wording decisions.The assistants were used for writing support, not to make the paper’s research or interpretation decisions.Source: G Use of AI Assistants · p. 17
  • Poetry fine-tuning produced seed-dependent changes on idioms and proverbs, with tables covering 15 unstable idioms and 7 unstable proverbs.The captions define unstable items as improved in some seeds and regressed in others; the supplied passages do not provide the underlying counts.Source: Table 16 · p. 18, Table 17 · p. 18
  • Tables report improvement and regression breakdowns for Kinayat idioms, Jawaher proverbs, and AraDiCE-Culture under cultural, figurative, and poetry fine-tuning.They specify accuracy before and after fine-tuning, percentage-point changes with paired-bootstrap confidence intervals, and counts of improved and worsened predictions; the captions also identify exact McNemar tests.Source: Table 19 · p. 19, Table 20 · p. 20, Table 21 · p. 21

¯ ÈAªË@

  • The cultural fine-tuning appendix lists the most frequently improved and regressed idioms across models fine-tuned on ArabCulture and Palm.Table 22 describes a top-15 list; the supplied caption does not show its entries or counts.Source: Table 22 · p. 21
  • The appendix lists the most frequently improved and regressed proverbs across fine-tuned models, with counts and dialect varieties.Table 23 covers all fine-tuned models; Table 18 separately describes net fine-tuning effects by dialect on the proverb task, but the supplied passages do not expose its values.Source: Table 18 · p. 18, Table 23 · p. 22

àA¿ ñË ð Q J.ºË@ ÐC¿ ©ÖÞ @

  • Poetry fine-tuning tables list the most frequently improved and regressed idioms and proverbs across models, including proverb counts by dialect.Tables 25 and 26 specify FannOrFlop; Table 24 describes net poetry fine-tuning effects by dialect on the proverb task, without exposing values in the supplied passages.Source: Table 25 · p. 23, Table 24 · p. 23, Table 26 · p. 24
Loading 2608.18361v1…