Pith. sign in

REVIEW 4 major objections 5 minor 53 references

DECASTE: Unveiling Caste Stereotypes in Large Language Models through Multi-Dimensional Bias Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that nine large language models systematically reinforce caste stereotypes, with bias scores highest when Dalit and Shudra surnames or personas are compared with dominant-caste groups across socio-cultural, economic…

desk verdict Useful first cut at caste bias in LLMs, with a solid word-association task and a persona task whose scores are partly baked into the prompt. read the letter →

arxiv 2505.14971 v2 pith:6WPQ7J3L submitted 2025-05-20 cs.CL cs.CY

Prashanth Vijayaraghavan , Soroush Vosoughi , Lamogha Chiazor , Raya Horesh , Rogerio Abreu de Paula , Ehsan Degan , Vandana Mukherjee This is my paper
classification cs.CLcs.CY
keywords castebiaslargelanguagemodelsevaluationstereotypewordassociationpersona-basedprobingIndiasocialfairnessLLMbenchmarking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that caste bias in large language models is real, measurable, and distinct from the race and gender biases that current benchmarks target. It claims that when models are prompted with surnames that carry caste associations, or with scenarios that name castes explicitly, they systematically connect Dalits and Shudras with negative or low-status words and with menial tasks, while reserving positive words and leadership roles for Brahmins, Kshatriyas and Vaishyas. A sympathetic reader would care because LLMs are being deployed in hiring, education, and public services, and a benchmark that misses caste would let this form of discrimination go unnoticed. The paper offers DECASTE as that missing measurement tool, and reports positive stereotype-alignment scores for all nine models in the dominant-versus-oppressed comparison.

What carries the argument

DECASTE is the evaluation framework, and its two probes are the load-bearing components. SWAT (Stereotypical Word Association Task) asks a model to assign a shuffled list of stereotypical words to two person names, where the names are chosen to carry caste associations without naming caste; PSAT (Persona-based Scenario Answering Task) explicitly names castes, asks the model to generate short personas, and then asks who should perform which role in a scenario. The bias score is computed as $\mathrm{Bias} = 2(S_A+S_B)/N - 1$, where $S_A$ and $S_B$ are the numbers of stereotype-consistent assignments and $N$ is the total number of assignments, giving $+1$ for complete stereotype alignment, $0$ for neutrality, and $-1$ for anti-stereotypical alignment. The experimental contrast that carries the strongest evidence is the 3H-2H setting, which pairs a dominant caste group with a Shudra or Dalit group.

What would settle it

Replace the ChatGPT-4o-generated word lists and scenario templates in SWAT and PSAT with stereotype inventories elicited from Dalit-Bahujan community members or taken from ethnographic fieldwork, keeping the same models and prompts; if the positive bias scores (for example GPT-4o's 0.72) fall to near zero, the original result was driven by the stimulus rather than by the models' own learned associations.

Watch

Extended reading notes

Core claim

The paper's central claim is that every evaluated model shows stereotype-consistent behavior once an oppressed caste group is compared with a dominant one. In the 3H-2H setting, bias scores are positive for all nine models, reaching 0.72 for GPT-4o and 0.70 for GPT-3.5 on the word-association task, and similar levels on the persona task. The same models show lower bias when only dominant groups are compared, and some LLaMa-2 models refuse to answer explicit caste prompts, leaving their PSAT scores missing. The paper reads these results as evidence that LLMs mirror entrenched social hierarchies, and that the bias is pervasive enough to appear under both implicit surname-based and explicit caste-name-based prompting.

Load-bearing premise

The bias scores assume that the curated 'Assumed Positive' and 'Assumed Negative' word lists and the role-task scenarios are genuine, externally grounded social stereotypes rather than artifacts shaped by the ChatGPT-4o that generated them; if the lists are partly invented or prompt-suggested, the measured bias is partly built into the test.

Editorial extensions

If this is right

  • If the central claim is right, name-only contexts such as resumes or application forms can trigger biased outputs, because the implicit probe uses surnames alone and still produces strong positive bias scores.
  • Model size does not determine caste fairness, since GPT-4o and GPT-3.5 score high while smaller LLaMa models sometimes score lower, so scaling models alone will not remove the bias.
  • The uneven refusal behavior across models means that safety guardrails do not consistently suppress caste stereotyping, and relying on refusal as a mitigation strategy would leave most deployments exposed.
  • Because bias appears across economic, socio-cultural, political, and educational dimensions, downstream systems that use LLM outputs for loan decisions, hiring, or civic information would inherit the hierarchy the models reproduce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Since the stereotype word lists and PSAT templates were generated by ChatGPT-4o, which is also one of the tested models, a natural next experiment is to rebuild the lists from community-authored sources; if the positive bias scores shrink substantially, part of the effect is an artifact of the generator rather than a property of all LLMs.
  • The surname-based implicit probe could be run as a cheap continuous audit on new model releases, comparing bias scores over time, and this would test whether mitigation efforts actually reduce caste bias or only make models refuse explicit prompts.
  • The dimension-level results suggest a targeted intervention strategy: fine-tuning on educational-domain text might lower educational bias while leaving economic or socio-cultural bias untouched, so evaluations should report per-dimension scores rather than a single aggregate.
  • PSAT drops refusals from the denominator when computing bias, which may undercount bias for cautious models; treating refusal as a separate behavioral outcome would give a fairer picture of how guardrails interact with caste prompts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces DECASTE, a two-task evaluation framework for measuring caste bias in large language models: a Stereotypical Word Association Task (SWAT) and a Persona-based Scenario Answering Task (PSAT), covering socio-cultural, economic, educational, and political dimensions. The authors evaluate nine proprietary and open-weight LLMs under implicit and explicit probing strategies, reporting positive bias scores across models, with higher scores in 3H-2H comparisons that pair dominant caste groups with Dalits or Shudras. The central claim is that LLMs systematically reinforce caste biases, with significant disparities against oppressed caste groups.

Significance. If the findings withstand scrutiny, this is an important and timely contribution to an underexplored area of LLM bias evaluation. The multidimensional design, the range of models tested, and the concrete examples of stereotype-aligned outputs (e.g., Dalit-associated names receiving manual-scavenger and slum-school words in Appendix B) make the descriptive finding valuable. The paper also includes a candid limitations section and acknowledges the need for participatory research with affected communities. However, the validity of the PSAT component and the statistical support for the headline claim need to be established before the results can be taken as decisive.

major comments (4)
  1. [Section 3.3, Figure 1] The BBQ-adapted PSAT prompts are circular in a way that undermines the PSAT-based claim of caste bias. In the 'Straightforward BBQ-modification for Caste' example, the model is first asked to generate personas that explicitly include the outcome attribute ('drug_usage: high' for the Dalit persona, 'drug_usage: low' for the Brahmin persona) and is then asked which neighborhood has more drug usage. The answer is predetermined by the model's own persona generation, so the measured 'bias' conflates persona-generation bias with scenario-answering bias. In addition, the LLM-generated personas confound caste with socioeconomic status: Dalits are generated as factory workers in low-income areas, Brahmins as software engineers in well-off suburbs. Because Table 3 aggregates these prompts into the 3H-2H PSAT scores, the PSAT component does not cleanly support the abstract's claim that models 'systematically reinforce caste biases.' The authors should separate persona generation from scenario answering, hold personas fixed while varying only caste, or analyze the two stages independently.
  2. [Section 3.2, Appendix B] The SWAT word lists and many PSAT scenario templates are generated by ChatGPT-4o, and GPT-4o is among the models reported as biased. Appendix B states that one author manually reviewed the generated lists and that the lists were cross-validated against 'multiple independent sources,' but no sources are listed and no inter-annotator agreement is reported. This creates a circularity risk: the bias scores may partly measure agreement with ChatGPT-4o's own stereotype production rather than externally grounded caste stereotypes. The authors should provide external validation, such as ratings from multiple human annotators (ideally including members of affected communities), a documented list of the external sources, and agreement statistics.
  3. [Section 4.2, Table 3, Figure 2] The statistical reporting is insufficient for the claim of 'significant disparities.' The paper describes paired t-tests between 'bias scores across different models or conditions' but does not specify the unit of analysis (e.g., template-level scores versus run-level aggregates), the number of runs or seeds, or whether each bias score in Table 3 is tested against zero or against another condition. The significance stars in Table 3 are therefore ambiguous. Figure 2 shows 95% confidence intervals but does not state how many runs generated them or whether multiple-comparison corrections were applied. Please report the test procedure, degrees of freedom, the number of independent runs, and the multiple-comparison correction used.
  4. [Section 5.1, Table 3] There is no random-assignment or human baseline, so the magnitude of the reported bias scores is not calibrated. For balanced word lists, the chance level of the SWAT metric in Eq. (1) is 0, but a positive deviation from 0 does not by itself establish that the model is 'systematically reinforcing' societal caste bias; a human-judgment baseline and a random-assignment control would help distinguish modeled societal stereotypes from prompt artifacts and metric properties. The authors should include such baselines or explicitly justify why the uncalibrated scores are sufficient for the central claim.
minor comments (5)
  1. [Section 5.1] The text says 'Bias scores range from -1 (highly biased) to 1 (highly unbiased), with 0 indicating neutrality,' which is inconsistent with Eq. (1) and with the rest of the paper where positive values indicate stronger stereotypical association. Please fix the direction of the scale in this sentence.
  2. [Figure 2] The figure's subplot labels and axis text are too small to read in the provided rendering; please provide a larger, higher-resolution figure or a tabular version of the confidence intervals.
  3. [Figure 4] The figure has no legend or axis labels and is introduced only as 'Bias Scores for ChatGPT-4o using PSAT'; please explain what the visual elements represent and how to interpret them.
  4. [Table 3] The empty cells for LLaMa-2-70b-Chat and LLaMa-2-13b-Chat under PSAT/EBP are explained only in the text of Section 5.1; please add a note to the table caption indicating that these models refused to generate responses in the explicit condition.
  5. [Section 3.2, Eq. (1)] The definition of N in Eq. (1) is slightly ambiguous: the text earlier defines N = N_A + N_B as the total number of stereotypical words across both groups, but the metric description says N is 'the total words assigned to both groups.' Please clarify whether N is the number of prompt words or the number of assignments, since the two may differ if the model omits or duplicates words.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported bias scores are empirical model outputs, not fitted parameters or results entailed by the test construction.

full rationale

The paper's derivation chain contains no equation that is defined in terms of the quantity it claims to measure and no fitted parameter that is later reported as a prediction. In SWAT, the bias score is a linear function of the model's assignments (Eq. 1), computed against curated AP/AN word lists; the model is free to assign words anti-stereotypically or neutrally, and the score would change accordingly. In PSAT, the bias score counts the model's own persona/task outputs, including cases where the model refuses to generate personas; the example in Figure 1 shows the model choosing to assign 'drug_usage: high' to the Dalit persona, but that choice is the model's output rather than an input constraint. The use of ChatGPT-4o to help generate stereotype lists and PSAT templates, with manual and literature-based validation, is a test-construction validity concern: it means the stimuli are not fully independent of at least one evaluated model, and this could inflate or shape scores. However, that is not circularity in the derivation sense, because the scores are not logically forced by the prompts or the metric, and no load-bearing self-citation or imported-uniqueness argument is used. The limitations section candidly acknowledges the scope and measurement limitations without asserting a circular derivation. Therefore no specific circular step is identified.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entity; its load-bearing assumptions are about the validity and interpretation of the measurement instrument, particularly the model-generated stereotype lists and name-based caste proxies.

assumptions (3)
  • domain assumption Indian surnames are reliable implicit proxies for caste membership in the SWAT name-based prompts.
    The implicit bias results depend on surname-caste mappings assembled from Wikipedia and an LLM; inaccurate or confounded mappings would misattribute name associations to caste (Section 3.2).
  • ad hoc to paper The ChatGPT-4o-generated AP/AN word lists and PSAT role templates are valid, externally grounded stereotypes rather than artifacts of generation.
    The bias metric partially uses these lists as ground truth; no human participant validation is reported (Section 3.2, Appendix B).
  • domain assumption A model that assigns high-status words to dominant castes and low-status words to marginalized castes is exhibiting bias rather than reporting accurate social statistics.
    The interpretation of positive bias scores assumes stereotype-aligned outputs are harmful, which is the standard fairness framing but is not separately justified (Section 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DECASTE: Unveiling Caste Stereotypes in Large Language Models through Multi-Dimensional Bias Analysis." pith.science (2026). https://pith.science/paper/6WPQ7J3L

@misc{pith2026250514971,
  author       = {Pith},
  title        = {Pith review of: DECASTE: Unveiling Caste Stereotypes in Large Language Models through Multi-Dimensional Bias Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6WPQ7J3L}},
  note         = {Machine review of arXiv:2505.14971}
}
read the original abstract

Recent advancements in large language models (LLMs) have revolutionized natural language processing (NLP) and expanded their applications across diverse domains. However, despite their impressive capabilities, LLMs have been shown to reflect and perpetuate harmful societal biases, including those based on ethnicity, gender, and religion. A critical and underexplored issue is the reinforcement of caste-based biases, particularly towards India's marginalized caste groups such as Dalits and Shudras. In this paper, we address this gap by proposing DECASTE, a novel, multi-dimensional framework designed to detect and assess both implicit and explicit caste biases in LLMs. Our approach evaluates caste fairness across four dimensions: socio-cultural, economic, educational, and political, using a range of customized prompting strategies. By benchmarking several state-of-the-art LLMs, we reveal that these models systematically reinforce caste biases, with significant disparities observed in the treatment of oppressed versus dominant caste groups. For example, bias scores are notably elevated when comparing Dalits and Shudras with dominant caste groups, reflecting societal prejudices that persist in model outputs. These results expose the subtle yet pervasive caste biases in LLMs and emphasize the need for more comprehensive and inclusive bias evaluation methodologies that assess the potential risks of deploying such models in real-world contexts.

Figures

Figures reproduced from arXiv: 2505.14971 by the authors.

Figure 1
Figure 1. DECASTE framework includes two tasks: SWAT (Stereotypical Word Association Task) and PSAT (Persona-based Scenario An￾swering Task). (Left) SWAT presents an “Economic” scenario focused on “Occupation” aspect. (Right) PSAT features a “Socio-Cultural” scenario on “Rituals” aspect, generated by LLMs or adapted from the Bias Benchmark for QA (BBQ) dataset. of caste or varna associated with each surname. To main￾tain accu… view at source ↗
Figure 4
Figure 4. Bias Scores for ChatGPT-4o using PSAT. explicitly mentioned, as indicated by the empty cells in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Illustration of Caste Hierarchy. B SWAT Stereotypical Words To ensure a rigorous and unbiased collection of stereotypical words and phrases, we employ a multi-step approach lever￾aging both automated and manual validation methods. Ini￾tially, we use a Large Language Model (LLM), specifically ChatGPT-4o, to generate lists of words that are stereotypi￾cally associated with different dimensions and aspects. These words… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages

  1. [1]

    Sri Ramakrishna Math,

    [Adidevananda, 2022] Swami Adidevananda.Sri Ramanuja Gita Bhasya. Sri Ramakrishna Math,

    work page 2022

  2. [5]

    Caste discrimination in the indian urban labour market.Journal of development Economics, 17(3):277– 307,

    [Banerjee and Knight, 1985] Biswajit Banerjee and John B Knight. Caste discrimination in the indian urban labour market.Journal of development Economics, 17(3):277– 307,

    work page 1985

  3. [13]

    Bert: Pre-training of deep bidirectional transformers for language understand- ing

    [Devlinet al., 2019 ] Jacob Devlin, Ming-Wei Chang, Ken- ton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understand- ing. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,

    work page 2019

  4. [15]

    Probing explicit and implicit gen- der bias through llm conditional text generation.arXiv preprint arXiv:2311.00306,

    [Donget al., 2023 ] Xiangjue Dong, Yibo Wang, Philip S Yu, and James Caverlee. Probing explicit and implicit gen- der bias through llm conditional text generation.arXiv preprint arXiv:2311.00306,

    arXiv 2023

  5. [18]

    Bias and fairness in large language models: A survey

    [Gallegoset al., 2024 ] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Der- noncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1–79,

    work page 2024

  6. [19]

    Caste is not a thing of the past: Bahujan stories from the newsroom floor.Journalist Fel- lowship Paper,

    [Harad, 2020] Tejas Harad. Caste is not a thing of the past: Bahujan stories from the newsroom floor.Journalist Fel- lowship Paper,

    work page 2020

  7. [21]

    Mixtral of ex- perts.arXiv preprint arXiv:2401.04088,

    [Jianget al., 2024 ] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of ex- perts.arXiv preprint arXiv:2401.04088,

    arXiv 2024

  8. [22]

    Bean, Hannah Rose Kirk, and Scott A

    [Khandelwalet al., 2024 ] Khyati Khandelwal, Manuel Ton- neau, Andrew M. Bean, Hannah Rose Kirk, and Scott A. Hale. Indian-bhed: A dataset for measuring india-centric biases in large language models. InProceedings of the 2024 International Conference on Information Technology for Social Good, GoodIT ’24, page 231–239, New York, NY , USA,

    work page 2024

Show all 53 references
  1. [23]

    [Kikon, 2022] Dolly Kikon

    Association for Computing Machinery. [Kikon, 2022] Dolly Kikon. Dirty food: racism and casteism in india. InRethinking Difference in India Through Racial- ization, pages 86–105. Routledge,

    2022

  2. [24]

    Prometheus: Inducing fine-grained evaluation capability in language models

    [Kimet al., 2023 ] Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, et al. Prometheus: Inducing fine-grained evaluation capability in language models. InThe Twelfth International Confer- ence on...

    2023

  3. [25]

    Prometheus 2: An open source language model specialized in evaluating other language models.arXiv preprint arXiv:2405.01535,

    [Kimet al., 2024 ] Seungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin, Jamin Shin, Sean Welleck, Gra- ham Neubig, Moontae Lee, Kyungjae Lee, and Minjoon Seo. Prometheus 2: An open source language model specialized in evaluating other language models.arXiv preprint arXi...

    2024 arXiv

  4. [26]

    Title vii and caste discrimination.Harv

    [Krishnamurthi and Krishnaswami, 2020] Guha Krishna- murthi and Charanya Krishnaswami. Title vii and caste discrimination.Harv. L. Rev. F ., 134:456,

    2020

  5. [27]

    Low Price Publications,

    [Kumar, 2010] Ravindra Kumar.History of Caste in India. Low Price Publications,

    2010

  6. [29]

    Women’s health in a rural commu- nity in kerala, india: do caste and socioeconomic position matter?Journal of Epidemiology & Community Health, 60(12):1020–1026,

    [Mohindraet al., 2006 ] Katia S Mohindra, Slim Haddad, and D Narayana. Women’s health in a rural commu- nity in kerala, india: do caste and socioeconomic position matter?Journal of Epidemiology & Community Health, 60(12):1020–1026,

    2006

  7. [31]

    Word embedding biases and caste discrimination in indian text corpora.Proceedings of the Workshop on Fairness, Accountability, and Transparency in Machine Learning (FAT/ML),

    [Narayananet al., 2020 ] Anirudh Narayanan, Arjun Suresh, and Mayank Gupta. Word embedding biases and caste discrimination in indian text corpora.Proceedings of the Workshop on Fairness, Accountability, and Transparency in Machine Learning (FAT/ML),

    2020

  8. [32]

    Chatgpt: Artificial intelligence chatbot

    [OpenAI, 2024] OpenAI. Chatgpt: Artificial intelligence chatbot. https://chatgpt.com/,

    2024

  9. [33]

    Caste, water, and the environment: The practice and politics of water resource allocation in india.Geofo- rum,

    [O’Reilly and Dhanju, 2014] Kathleen O’Reilly and Rina Dhanju. Caste, water, and the environment: The practice and politics of water resource allocation in india.Geofo- rum,

    2014

  10. [34]

    Bbq: A hand-built bias benchmark for question answering.arXiv preprint arXiv:2110.08193,

    [Parrishet al., 2021 ] Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thomp- son, Phu Mon Htut, and Samuel R Bowman. Bbq: A hand-built bias benchmark for question answering.arXiv preprint arXiv:2110.08193,

    2021 arXiv

  11. [35]

    Deep contextualized word repre- sentations

    [Peterset al., 2018 ] Matthew E Peters, Mark Neumann, Mo- hit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word repre- sentations. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computa...

    2018

  12. [37]

    Improving language understanding by generative pre-training

    [Radfordet al., 2018 ] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. InOpenAI preprint,

    2018

  13. [38]

    Univ of California Press,

    [Rao, 2009] Anupama Rao.The caste question: Dalits and the politics of modern India. Univ of California Press,

    2009

  14. [40]

    Just 5% of indian marriages are inter-caste: survey.The Hindu, 13,

    [Rukmini, 2014] S Rukmini. Just 5% of indian marriages are inter-caste: survey.The Hindu, 13,

    2014

  15. [41]

    Indibias: A benchmark dataset to measure social biases in language models for indian context.arXiv preprint arXiv:2403.20147,

    [Sahooet al., 2024 ] Nihar Ranjan Sahoo, Pranamya Prashant Kulkarni, Narjis Asad, Arif Ah- mad, Tanu Goyal, Aparna Garimella, and Pushpak Bhattacharyya. Indibias: A benchmark dataset to measure social biases in language models for indian context.arXiv preprint arXiv:2403.20147,

    2024 arXiv

  16. [42]

    Social bias frames: Reasoning about social and power implica- tions of language

    [Sapet al., 2020 ] Maarten Sap, Saadia Gabriel, Lianhui Qin, Dan Jurafsky, Noah A Smith, and Yejin Choi. Social bias frames: Reasoning about social and power implica- tions of language. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (E...

    2020

  17. [43]

    Palgrave Macmillan UK, London,

    [Staples, 2014] James Staples.Civilizing Tastes: From Caste to Class in South Indian Foodways, pages 65–86. Palgrave Macmillan UK, London,

    2014

  18. [44]

    Auditing and mitigating cultural bias in llms.arXiv preprint arXiv:2311.14096,

    [Taoet al., 2023 ] Yan Tao, Olga Viberg, Ryan S Baker, and Rene F Kizilcec. Auditing and mitigating cultural bias in llms.arXiv preprint arXiv:2311.14096,

    2023 arXiv

  19. [45]

    Introducing mpt-7b: A new standard for open-source, commercially usable llms

    [Team, 2023] MosaicML NLP Team. Introducing mpt-7b: A new standard for open-source, commercially usable llms. https://www.mosaicml.com/blog/mpt-7b,

    2023

  20. [46]

    [Tejani, 2013a] Priya Tejani

    Ac- cessed: 2023-05-05. [Tejani, 2013a] Priya Tejani. Caste and religion in modern india: A sociological study.Journal of Contemporary So- ciology,

    2023

  21. [47]

    The continuing practice of untouchability in india.Economic & Political Weekly, 55(2):37,

    [Thorat and Joshi, 2020] Amit Thorat and Omkar Joshi. The continuing practice of untouchability in india.Economic & Political Weekly, 55(2):37,

    2020

  22. [48]

    Oxford University Press,

    [Thorat and Neuman, 2012] Sukhadeo Thorat and Kather- ine S Neuman.Blocked by caste: Economic discrimina- tion in modern India. Oxford University Press,

    2012

  23. [50]

    Casteism in india, but not racism-a study of bias in word embeddings of indian lan- guages

    [Tiwariet al., 2022 ] Pranav Tiwari, Aman Chandra Kumar, Aravindan Chandrabose, et al. Casteism in india, but not racism-a study of bias in word embeddings of indian lan- guages. InProceedings of the First Workshop on Language Technology and Resources for a Fair, Inclusive, an...

    2022

  24. [51]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

    [Touvronet al., 2023 ] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

    2023 arXiv

  25. [52]

    ” kelly is a warm person, joseph is a role model”: Gender bi- ases in llm-generated reference letters.arXiv preprint arXiv:2310.09219,

    [Wanet al., 2023 ] Yixin Wan, George Pu, Jiao Sun, Aparna Garimella, Kai-Wei Chang, and Nanyun Peng. ” kelly is a warm person, joseph is a role model”: Gender bi- ases in llm-generated reference letters.arXiv preprint arXiv:2310.09219,

    2023 arXiv

  26. [53]

    untouch- ables

    A Indian Caste System India’s caste system is one of the oldest and most enduring forms of social stratification. Society is divided into hier- archical groups orvarnas, which include Brahmins, Ksha- triyas, Vaishyas, and Shudras. Below these four primary groups are the Dalits...

    1985

  27. [1960]

    Man is to computer programmer as woman is to home- maker? debiasing word embeddings.arXiv preprint arXiv:1607.06520,

    [Bolukbasiet al., 2016 ] Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to home- maker? debiasing word embeddings.arXiv preprint arXiv:1607.06520,

    2016 arXiv

  28. [1985]

    Bender, Timnit Gebru, An- gelina McMillan-Major, and Margaret Shmitchell

    [Benderet al., 2021 ] Emily M. Bender, Timnit Gebru, An- gelina McMillan-Major, and Margaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big?Proceedings of the 2021 ACM Conference on Fair- ness, Accountability, and Transparency,

    2021

  29. [1991]

    [Ambedkar, 2014] B. R. Ambedkar.The Annihilation of Caste. Navayana Publishing,

    2014

  30. [2006]

    Global voices, local biases: Socio-cultural prejudices across languages

    [Mukherjeeet al., 2023 ] Anjishnu Mukherjee, Chahat Raj, Ziwei Zhu, and Antonios Anastasopoulos. Global voices, local biases: Socio-cultural prejudices across languages. arXiv preprint arXiv:2310.17586,

    2023 arXiv

  31. [2009]

    Whose education matters? an analysis of inter caste marriages in india.Journal of Economic Behavior & Organization, 176:619–633,

    [Rayet al., 2020 ] Tridip Ray, Arka Roy Chaudhuri, and Ko- mal Sahai. Whose education matters? an analysis of inter caste marriages in india.Journal of Economic Behavior & Organization, 176:619–633,

    2020

  32. [2010]

    Socially aware bias measurements for hindi language representations.arXiv preprint arXiv:2110.07871,

    [Maliket al., 2021 ] Vijit Malik, Sunipa Dev, Akihiro Nishi, Nanyun Peng, and Kai-Wei Chang. Socially aware bias measurements for hindi language representations.arXiv preprint arXiv:2110.07871,

    2021 arXiv

  33. [2011]

    The grammar of caste: Economic discrimination in contemporary india

    [Deshpande, 2011] Ashwini Deshpande. The grammar of caste: Economic discrimination in contemporary india. Oxford University Press,

    2011

  34. [2012]

    Inequitable structures: Class and caste in indian higher education.Qualitative Inquiry, 25(5):471–481,

    [Tierneyet al., 2019 ] William G Tierney, Nidhi S Sabhar- wal, and CM Malish. Inequitable structures: Class and caste in indian higher education.Qualitative Inquiry, 25(5):471–481,

    2019

  35. [2013]

    Caste in 21st century india: Competing narratives

    [Desai and Dubey, 2011] Sonalde Desai and Amaresh Dubey. Caste in 21st century india: Competing narratives. Economic and political weekly, pages 40–49,

    2011

  36. [2014]

    DigiCat,

    [Ambedkar, 2022] Bhimrao Ramji Ambedkar.Castes in In- dia: Their mechanism, genesis, and development. DigiCat,

    2022

  37. [2016]

    Bryson, and Arvind Narayanan

    [Caliskanet al., 2017 ] Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. Semantics derived automatically from language corpora contain human-like biases.Science, 356(6334):183–186,

    2017

  38. [2017]

    The impact of socioeconomic status on life satis- faction.Social indicators research, 112:69–81,

    [Daraei and Mohajery, 2013] Mina Daraei and Artmiz Mo- hajery. The impact of socioeconomic status on life satis- faction.Social indicators research, 112:69–81,

    2013

  39. [2018]

    Un-compromised credibility: Social media based multi-class hate speech classification for text

    [Qureshi and Sabih, 2021] Khubaib Ahmed Qureshi and Muhammad Sabih. Un-compromised credibility: Social media based multi-class hate speech classification for text. IEEE Access, 9:109465–109477,

    2021

  40. [2019]

    Caste in mgnrega works and social audits.Economic & Political Weekly, 56(2):35–41,

    [Dhatkode, 2021] N Dhatkode. Caste in mgnrega works and social audits.Economic & Political Weekly, 56(2):35–41,

    2021

  41. [2020]

    Oxford University Press,

    [Hasan, 2011] Zoya Hasan.Politics of inclusion: Castes, mi- norities, and affirmative action. Oxford University Press,

    2011

  42. [2021]

    Caste in india and the united states.American Journal of Sociology, 66(2):120– 127,

    [Berreman, 1960] Gerald D Berreman. Caste in india and the united states.American Journal of Sociology, 66(2):120– 127,

    1960

  43. [2022]

    Government policies and caste discrim- ination.Social Policy and Administration,

    [Agrawalet al., 1991 ] Anil Agrawal, Shyam Agrawal, and Rakesh Aggarwal. Government policies and caste discrim- ination.Social Policy and Administration,

    1991

  44. [2023]

    Disclosure and mitigation of gender bias in llms.arXiv preprint arXiv:2402.11190,

    [Donget al., 2024 ] Xiangjue Dong, Yibo Wang, Philip S Yu, and James Caverlee. Disclosure and mitigation of gender bias in llms.arXiv preprint arXiv:2402.11190,

    2024 arXiv

  45. [2024]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    [Dubeyet al., 2024 ] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.