Pith. sign in

REVIEW 5 major objections 6 minor 46 references

Exploring Gender Bias Beyond Occupational Titles

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Language-model gender bias is distributed across verbs, nouns, and occupations, and a new score can isolate each component's contribution.

desk verdict The GenderLexicon dataset is a real contribution, but the paper's headline claim—that action verbs are more male-biased than occupations—is an artifact of the unvalidated (1 - sim(g,c)) exponent, not a property of the raw language model. read the letter →

arxiv 2507.02679 v2 pith:L4DF73ZU submitted 2025-07-03 cs.CL

Ahmed Sabir , Rajesh Sharma This is my paper
classification cs.CL
keywords genderbiaslargelanguagemodelsclozeprobabilitycontextualoccupationalstereotypesamplificationwordembeddingscoreferenceresolution
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

This paper argues that gender bias in large language models is not confined to occupational stereotypes, and that action verbs and object nouns independently influence which pronoun a model chooses. To show this, the authors introduce a new template dataset, GenderLexicon, and a scoring metric, the ClozeGender Score, that amplifies a model's cloze probability for a pronoun by a factor derived from the semantic similarity between the pronoun and the sentence context. Across five datasets, including WinoBias, WinoGender, and a Japanese stereotype set, they report that object nouns bias toward female pronouns while action verbs bias toward male pronouns, with verbs showing a stronger male bias than occupations. The work matters because existing benchmarks built on occupation templates may substantially understate contextual bias, and the proposed score offers a per-component, explainable bias quantification.

What carries the argument

The load-bearing machinery is the ClozeGender Score, CGS(S) = P([cloze]:g | [w1,...,w_{n-1}])^(1 - sim(g,c)), which combines an LLM's cloze probability with a static word-embedding similarity so that a pronoun semantically close to the context word receives an amplified bias score. The second load-bearing component is the GenderLexicon dataset itself: 2,511 template sentences (837 unique) generated by ChatGPT-3.5 and manually corrected by human annotators, each containing an occupation, an object noun, an action verb, and a final pronoun slot, plus a gender-neutral variant that omits the occupation. The design ensures a single context word can be isolated as the bias source, and the score is computed over static embeddings (word2vec, GloVe, fastText, GN-GloVe, DD-GloVe) to keep the similarity source independent of the LLM being tested.

What would settle it

Compute the bias ratio for GenderLexicon with the raw cloze probability P(g|context) instead of CGS, and compare the ranking of action verbs versus occupations; if the ranking reverses or disappears, the paper's central finding depends on the unvalidated exponent rather than on the language model itself.

Watch

Extended reading notes

Core claim

The central claim of the paper is that gender bias in language models is a distributed contextual effect: occupation, object noun, and action verb each carry an independent gender signal, and the signal is not reducible to the occupation stereotype. The authors define the ClozeGender Score as CGS(S) = P(g | context)^(1 - sim(g, c)), where the raw cloze probability is raised to a power controlled by the semantic similarity between the pronoun g and the context word c, and they validate it on their GenderLexicon dataset plus WinoBias, WinoGender, CrowS-Pairs, and a newly constructed Japanese cultural stereotype dataset. Their key empirical result is that action verbs are more male-biased than occupations in both WinoBias and WinoGender, while object nouns are more female-biased in GenderLexicon, and that substituting a gender-neutral someone for the occupation shifts the bias direction and lowers KL divergence. They also show the same machinery measures occupational bias against marginalized ethnic groups, and they train a Race-Neutral GloVe embedding that reduces that bias.

Load-bearing premise

The load-bearing premise is that raising the model's pronoun probability to the power (1 minus the semantic similarity between pronoun and context) genuinely amplifies bias, so if that exponent is miscalibrated the measured bias scores and rankings could be artifacts of the similarity function rather than properties of the language model.

Editorial extensions

If this is right

  • Bias benchmarks built only on occupation templates, such as WinoBias and WinoGender, understate LLM gender bias because action verbs and object nouns carry independent bias signals that those templates do not isolate.
  • The ClozeGender Score gives a single number for how much a specific context word biases pronoun choice, which makes bias attribution explainable without requiring prompt probing of each model.
  • Debiasing that only neutralises occupation terms (e.g., GN-GloVe) leaves verb-driven male bias intact; the results imply debiasing pipelines must also target action verbs and object nouns.
  • The framework transfers across languages and social categories: it measured cultural stereotype bias in Japanese and occupational-racial bias toward marginalized groups, suggesting it is a general contextual-bias estimator rather than a gender-only score.

Reading between the lines

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

  • A direct testable extension would be to replace the static-embedding similarity in the exponent with a contextual similarity from a masked LM; if the verb-versus-occupation ranking is robust to that change, the finding is a property of the cloze distribution, not of the choice of similarity source.
  • The human evaluation suggests humans themselves inject male bias for unfamiliar occupations, so the CGS score may be closer to a fairness target than raw human judgment is; the paper does not make this normative claim itself.
  • Because the exponent amplifies any semantically similar pronoun-context pair, the score may overstate bias when the embedding similarity is high for spurious reasons (e.g., co-occurrence in biased corpora); a counterfactual swap of the object noun with a neutral one would reveal this.
  • The authors' binary gender framing leaves non-binary and genderless-language cases unaddressed, though the Japanese experiment with context-dependent pronouns hints at how the method could be adapted to languages where gender is not grammaticalized.
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

5 major / 6 minor

Summary. The paper argues that gender bias in language models extends beyond occupational stereotypes. It introduces a new dataset, GenderLexicon, with template sentences that link pronouns to action verbs, object nouns, and occupations, and proposes a ClozeGender Score (CGS) that adjusts raw cloze probabilities by raising them to an exponent 1 - sim(g, c), where sim(g, c) is a static word-embedding similarity between a pronoun and a context word. The authors evaluate CGS on GenderLexicon, WinoBias, WinoGender, CrowS-Pairs, and a new Japanese stereotype dataset, comparing GPT-2XL, BERT, ChatGPT, GPT-4 family, and several open-weight LLMs. The central reported findings are that object nouns are more female-biased, action verbs are more male-biased than occupations, and that occupation removal shifts bias direction.

Significance. The proposed dataset and the multilingual extension are potentially useful resources for fairness evaluation, since existing benchmarks are largely occupation-centric. The paper also compares a wide range of models and reports both intrinsic (WEAT, KL) and extrinsic measures. The GitHub release makes the dataset accessible for reproduction. However, the core quantitative claim 'action verbs are more male-biased than occupations' is not supported by the raw model probabilities: in Table 4, the unaugmented GPT-2XL row shows nearly identical male ratios for verbs and occupations in WinoBias (0.77 vs 0.76), and the large separation appears only after applying the unvalidated exponent in CGS. Because no derivation, ablation, or validation of the exponent is provided, the paper currently establishes a dataset contribution but not a validated bias-measurement contribution.

major comments (5)
  1. [§3, Eq. (1)] The exponent 1 - sim(g, c) is asserted without derivation or independent validation. Since sim(g, c) is computed separately for each gender pronoun, raising P(g|context) to different exponents for 'him' and 'her' can reorder male/female ratios even when the underlying LM probabilities are identical: a context word embedded closer to 'him' than to 'her' has its male exponent pushed toward zero, inflating the male share purely through the embedding geometry. The paper does not report the sim(g, c) values, does not include an ablation with a fixed exponent or with sim = 0, and does not test stability under monotonic transformations of the exponent. This is load-bearing because the abstract and Section 5 base the 'beyond occupational stereotypes' claim on the CGS-adjusted numbers.
  2. [§5, Table 4] The headline finding that action verbs are more male-biased than occupations is not present in the raw cloze probabilities. In the unaugmented GPT-2XL row for WinoBias, the male ratios are 0.77 for verbs and 0.76 for occupations; the reported separation (e.g., 0.93 vs 0.70 with the GloVe adjustment) is introduced entirely by the CGS exponent. The manuscript should report raw cloze ratios alongside adjusted ones and provide a direct statistical comparison of verb vs occupation male bias with confidence intervals. As written, the abstract's statement that the findings 'confirm the existence of gender biases beyond occupational stereotypes' overstates what the data show.
  3. [§6, Figure 1] The human evaluation relies on three naive subjects (HS2-HS4) and one professional annotation (HS1). No inter-annotator agreement, standard error, or significance test is reported, so the claims that 'humans introduce their own biases' and that the score 'reflects human bias' are not quantitatively established. This is not a minor issue because Section 9 uses 'reflects human bias' as the criterion for preferring ClozeLast over ClozeAll.
  4. [§2.2 and Limitations] The Limitations section acknowledges dependence on static word embeddings but asserts this is 'not a limitation in our case' because likelihood revision requires external information. This does not address the circularity concern: static embeddings are known to encode gender stereotypes, so using sim(g,c) as an exponent amplifies LM probabilities in the direction of embedding stereotypes, conflating the LM's contextual prediction with the embedding's static association. A concrete check would be to recompute all headline comparisons with the exponent fixed at 1 (raw cloze ratios) and with debiased embeddings, and to report the correlation between raw and adjusted ratios. The ablations in Section 9 vary cloze position and neutral entities but never vary the similarity mechanism itself.
  5. [Tables 2 and 4] The results vary substantially across embedding choices, and no error bars or significance tests are provided. For example, in WinoGender the occupation male ratio is 0.67 with GloVe but 0.49 with fasttext in Table 4, and the verb direction differs across embeddings in Table 2. Without uncertainty quantification, it is impossible to determine whether the verb/occupation ordering is statistically reliable or merely a function of the chosen embedding.
minor comments (6)
  1. [Tables 2-4] Several cells concatenate values without separators (e.g., '0.470.53', '0.40 0.60'), and some header rows are repeated or ambiguous. Please reformat all tables into clearly separated numeric columns with explicit row labels.
  2. [References] The citation 'Pennington, 2014' is incomplete; the GloVe paper should be cited with full author and publication details.
  3. [§4, WEAT formula] The displayed WEAT formula omits the standard deviation denominator of the standard effect-size formulation; as written, the reported WEAT values are not comparable to the usual WEAT effect size.
  4. [§7, CrowS-Pairs] The description of using GPT-4 to extract stereotype concepts and then 'manually corrected and revised all the instances' lacks detail on how many instances were corrected and whether the manual revisions were checked by multiple annotators; this should be reported for reproducibility.
  5. [§6, Figure 1] The relationship between HS1 (professional annotation) and HS2-HS4 (random subjects) is unclear from the figure and the prose; please define the subject groups and the exact evaluation task in the caption or main text.
  6. [§2.2, dataset counts] The text says the dataset contains 2,511 templates, 837 of which are unique, but later descriptions refer to '837 unique' without specifying whether this counts templates, sentences, or contexts; please reconcile the counts and definitions.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline verb-vs-occupation bias contrast is produced by the unvalidated 1-sim(g,c) exponent rather than by the LM's cloze probabilities, so the central empirical claim is partly circular.

  1. self definitional [Section 3 (ClozeGender equation) and Section 5, Table 4 discussion (WinoBias/WinoGender)]
    "ClozeGender Score (CGS) can be computed as: CGS(S) = P([cloze]:g|[w1, w2, . . . , wn−1])1−sim(g,c) ... When g and c are highly similar (with sim(g, c) close to 1), there is a relationship between gender and context (e.g. occupation), so the bias probability should be amplified ... Table 4 shows that our model interprets that the action verbs are more male-biased than occupations. GPT-2XL 0.77 0.22 0.76 0.23"

    Raising P(g|context) to the exponent 1-sim(g,c) is not a neutral calibration: for probabilities below 1, a smaller exponent produces a larger value, so whichever pronoun has higher embedding similarity to the context word gets a boosted probability. The sim(g,c) values come from static word embeddings (word2vec, GloVe, fasttext), which the paper itself measures as gender-biased via WEAT. Thus the CGS values in Table 4 encode embedding gender associations a second time, through the exponent, on top of the LM probabilities. The raw GPT-2XL row shows nearly identical male ratios for verbs and occupations (0.77 vs 0.76), so the reported verb > occupation finding is generated by the exponent and the similarity function rather than by the LM's cloze predictions.

full rationale

The paper does not rely on a load-bearing self-citation chain: citations to the authors' other work are peripheral, and no uniqueness theorem is invoked. The new GenderLexicon dataset, raw cloze probabilities, and human comparisons provide independent grounding for the existence of some gender bias. However, the paper's headline quantitative claim, that action verbs are more male-biased than occupations, does not appear in the raw LM probabilities; it emerges only after the 1-sim(g,c) exponentiation in the ClozeGender Score. Because static embeddings already encode gender associations (as the paper's own WEAT measurements show), the exponentiation imports those associations into the reported 'amplified bias' and can reorder the verb/occupation comparison by construction. This makes the central finding partially circular, while the raw probabilities and dataset retain independent content, which is why the score is 6 rather than higher.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The method depends on several unvalidated domain assumptions: cloze probability as a bias measure, embedding cosine similarity as a proxy for gender-context association, and a binary gender setup. No free parameters are fitted; the main risk is the ad hoc exponent in CGS.

assumptions (4)
  • domain assumption Cloze probability of a pronoun given a sentence predicts implicit gender bias.
    The method interprets P([cloze]:g|context) as the initial bias (Section 3).
  • domain assumption Cosine similarity between a pronoun and a context word in static word embeddings is a valid measure of semantic association for bias amplification.
    Used in the exponent 1 - sim(g,c) in CGS; not independently validated (Section 3).
  • standard math The amplification formula b(co,g) from Zhao et al. (2017) is appropriate for this score.
    Borrowed directly from prior work; assumed to transfer (Section 4).
  • domain assumption Gender is treated as binary in all evaluations.
    The authors acknowledge in the Ethics Statement that this oversimplifies gender.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Gender Bias Beyond Occupational Titles." pith.science (2026). https://pith.science/paper/L4DF73ZU

@misc{pith2026250702679,
  author       = {Pith},
  title        = {Pith review of: Exploring Gender Bias Beyond Occupational Titles},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L4DF73ZU}},
  note         = {Machine review of arXiv:2507.02679}
}
read the original abstract

In this work, we investigate the correlation between gender and contextual biases, focusing on elements such as action verbs, object nouns, and particularly on occupations. We introduce a novel dataset, GenderLexicon, and a framework that can estimate contextual bias and its related gender bias. Our model can interpret the bias with a score and thus improve the explainability of gender bias. Also, our findings confirm the existence of gender biases beyond occupational stereotypes. To validate our approach and demonstrate its effectiveness, we conduct evaluations on five diverse datasets, including a Japanese dataset.

Figures

Figures reproduced from arXiv: 2507.02679 by the authors.

Figure 1
Figure 1. Human evaluation. We compare the profes [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 23 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    AI@Meta . 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card . GitHub Repository

    work page 2024

  3. [3]

    Haozhe An, Xiaojiang Liu, and Donald Zhang. 2022. Learning bias-reduced word embeddings using dictionary definitions. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1139--1152

    work page 2022

  4. [4]

    Alex Andonian. 2023. https://doi.org/10.5281/zenodo.5879544 GPT-NeoX: Large Scale Autoregressive Language Modeling in PyTorch

    doi:10.5281/zenodo.5879544 2023

  5. [5]

    Bailey, Adina Williams, and Andrei Cimpian

    April H. Bailey, Adina Williams, and Andrei Cimpian. 2022. https://doi.org/10.1126/sciadv.abm2463 Based on billions of words on the internet, people = men . Science Advances, 8(13)

    work page doi:10.1126/sciadv.abm2463 2022

  6. [6]

    Jayadev Bhaskaran and Isha Bhallamudi. 2019. Good secretaries, bad truck drivers? occupational gender stereotypes in sentiment analysis. arXiv preprint arXiv:1906.10256

    work page Pith review arXiv 2019

  7. [7]

    Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In ACL

    work page 2021

  8. [8]

    Sergey V Blok, Douglas L Medin, and Daniel N Osherson. 2007. Induction as conditional probability judgment. Memory & Cognition

    work page 2007

Show all 46 references
  1. [9]

    Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. TACL

    2017

  2. [10]

    Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science

    2017

  3. [11]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Kaijie Zhu, Hao Chen, Linyi Yang, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. 2023. A survey on evaluation of large language models. arXiv preprint arXiv:2307.03109

    2023 arXiv

  4. [12]

    DeepSeek-AI. 2025. http://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    2025 arXiv

  5. [13]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT

    2019

  6. [14]

    Virginia K Felkner, Ho-Chun Herbert Chang, Eugene Jang, and Jonathan May. 2023. Winoqueer: A community-in-the-loop benchmark for anti-lgbtq+ bias in large language models. arXiv preprint arXiv:2306.15087

    2023 arXiv

  7. [15]

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. 2023. Bias and fairness in large language models: A survey. arXiv preprint arXiv:2309.00770

    2023 arXiv

  8. [16]

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462

    2020 arXiv

  9. [17]

    Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Mu \ n oz S \'a nchez, Mugdha Pandya, and Adam Lopez. 2020. Intrinsic bias metrics do not correlate with application bias. arXiv preprint arXiv:2012.15859

    2020 arXiv

  10. [18]

    Monica Gonzalez-Marquez. 2007. Methods in cognitive linguistics, volume 18. John Benjamins Publishing

    2007

  11. [19]

    Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. 2018. Learning word vectors for 157 languages. In LREC

    2018

  12. [20]

    Ignacio Iacobacci. 2015. Sensembed: Learning sense embeddings for word and relational similarity. In ACL

    2015

  13. [21]

    Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala, and Naoaki Okazaki. 2022. Gender bias in masked language models for multiple languages. arXiv preprint arXiv:2205.00551

    2022 arXiv

  14. [22]

    Elisabeth Kaukonen, Ahmed Sabir, and Rajesh Sharma. 2025. How aunt-like are you? exploring gender bias in the genderless estonian language: A case study. In NoDaLiDa/Baltic-HLT 2025

    2025

  15. [23]

    Rafal Kocielnik, Shrimai Prabhumoye, Vivian Zhang, Roy Jiang, R Michael Alvarez, and Anima Anandkumar. 2023. Biastestgpt: Using chatgpt for social bias testing of language models. arXiv preprint arXiv:2302.07371

    2023 arXiv

  16. [24]

    Hadas Kotek, Rikker Dockum, and David Q. Sun. 2023. https://arxiv.org/abs/2308.14921 Gender bias in llms

    2023 arXiv

  17. [25]

    Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning

    2012

  18. [26]

    Aaron Levine, Connie Huang, Chenguang Wang, Eduardo Batista, Ewa Szymanska, Hongyi Ding, Hou Wei Chou, Jean-Fran c ois Pessiot, Johanes Effendi, Justin Chiu, et al. 2024. Rakutenai-7b: Extending large language models for japanese. arXiv e-prints, pages arXiv--2403

    2024

  19. [27]

    Shahar Levy, Koren Lazar, and Gabriel Stanovsky. 2021. Collecting a large-scale gender bias dataset for coreference resolution and machine translation. arXiv preprint arXiv:2109.03858

    2021 arXiv

  20. [28]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In NeurIPS

    2013

  21. [29]

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel Bowman. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In EMNLP

    2020

  22. [30]

    OpenAI. 2022. https://openai.com/blog/chatgpt Chatgpt: Optimizing language models for dialogue

    2022

  23. [31]

    OpenAI. 2024. Gpt-4o system card. https://openai.com/index/gpt-4o-system-card/. Accessed: 2024-08-08

    2024

  24. [32]

    Jeffrey Pennington. 2014. Glove: Global vectors for word representation. EMNLP

    2014

  25. [33]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog

    2019

  26. [34]

    Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme . 2018. Gender bias in coreference resolution. In NAACL

    2018

  27. [35]

    Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. arXiv preprint arXiv:1909.01326

    2019 arXiv

  28. [36]

    Smith, and Luke Zettlemoyer

    Gabriel Stanovsky, Noah A. Smith, and Luke Zettlemoyer. 2019. https://doi.org/10.18653/v1/P19-1164 Evaluating gender bias in machine translation . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1679--1684, Florence, Italy. Ass...

    2019 doi

  29. [37]

    Wilson L Taylor. 1953. Cloze procedure: A new tool for measuring readability. Journalism quarterly, 30(4):415--433

    1953

  30. [38]

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. In NeurIPS

    2020

  31. [39]

    Yixin Wan and Kai-Wei Chang. 2024. White men lead, black women help: Uncovering gender, racial, and intersectional bias in language agency. arXiv preprint arXiv:2404.10508

    2024 arXiv

  32. [40]

    Ben Wang and Aran Komatsuzaki. 2021. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model . https://github.com/kingoflolz/mesh-transformer-jax

    2021

  33. [41]

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. arXiv preprint arXiv:1707.09457

    2017 arXiv

  34. [42]

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. arXiv preprint arXiv:1804.06876

    2018 arXiv

  35. [43]

    Jinman Zhao, Yitian Ding, Chen Jia, Yining Wang, and Zifan Qian. 2024. Gender bias in large language models across multiple languages. arXiv preprint arXiv:2403.00277

    2024 arXiv

  36. [44]

    Yi Zhou, Masahiro Kaneko, and Danushka Bollegala. 2022. Sense embeddings are also biased--evaluating social biases in static and contextualised sense embeddings. arXiv preprint arXiv:2203.07523

    2022 arXiv

  37. [45]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.