Pith. sign in

REVIEW 5 major objections 6 minor 24 references

CMLFormer: A Dual Decoder Transformer with Switching Point Learning for Code-Mixed Language Modeling

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A dual-decoder Transformer with switching-point and translation pretraining lifts Hinglish hate-speech F1 by 0.18 over a same-data BERT baseline.

desk verdict New dual-decoder architecture for code-mixed text, but the paper's own Table 1 undercuts the SPP attribution in the conclusion. read the letter →

arxiv 2505.12587 v1 pith:QDT6BMME submitted 2025-05-19 cs.CL cs.LG

Aditeya Baral , Allen George Ajith , Roshan Nayak , Mrityunjay Abhijeet Bhanja This is my paper · ORCID
classification cs.CLcs.LG
keywords code-mixedlanguagemodelingswitchingpointsdual-decodertransformerHinglishmultitaskpretraininghatespeechdetectioncross-lingualrepresentationlearningHASOC2021
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

CMLFormer is designed to test a specific hypothesis: a language model pretrained to know where language switches occur in code-mixed text, and to translate those mixed sentences into each constituent language, will represent code-mixed input better than models pretrained only on generic masked-language modeling. The paper builds a Transformer with one shared encoder and two synchronized decoders, one for the base language and one for the mixing language, and trains it on a small augmented Hinglish corpus with switching-point labels and parallel Hindi and English translations. After pretraining, only the encoder is fine-tuned on the HASOC-2021 Hinglish hate-speech benchmark, where the authors report F1 gains of 0.18 and 0.05 over a same-data BERT-base baseline from the translation and switching-point objectives, respectively. If this holds, explicit structural supervision may be a viable route to code-mixed NLP without relying on larger models or larger monolingual corpora.

What carries the argument

The load-bearing mechanism is the synchronized dual-decoder cross-attention sub-layer. Two Transformer decoders share one encoder; after each decoder's encoder-decoder attention, the base-language decoder attends to the mixing-language decoder's current hidden states and vice versa, with learned linear projections mapping each decoder's features into the other's space. This couples generation of the Hindi and English translations of the same code-mixed sentence, forcing the encoder to produce representations aligned across both languages. The other mechanism is switching-point prediction: a binary token-level classification head on the encoder, trained with word-level switch labels aligned to subwords, that teaches the encoder where language transitions occur.

What would settle it

Retrain CMLFormer on a human-annotated subset of the same 10,000 augmented Hinglish sentences, with manually verified switching-point labels and translations, and compare HASOC-2021 F1 against the model trained on LLM-generated labels; if the human-labeled model does not match or exceed the reported F1 gains, the claimed gains depend on the LLM labeling distribution rather than on switching structure.

Watch

Extended reading notes

Core claim

The central discovery claimed is that explicit structural supervision specific to code-mixing—per-token switching-point prediction and bilingual translation into both constituent languages—can teach a shared encoder representations that transfer better to downstream code-mixed classification than masked-language modeling alone. In the reported experiments, the translation objective (BiLTM) added to MLM raises HASOC-2021 hate-speech F1 from 0.249 for a same-data BERT-base to 0.431; the combination of MLM, BiLTM, and switching-point prediction reaches 0.295, which the paper counts as an absolute F1 gain of 0.05 over the baseline. Attention analysis shows the pretrained encoder allocates high attention around switching points regardless of switch direction or position, whereas BERT does not. The paper also reports that adding further objectives (BTSP, TLC, and CMI) degrades downstream performance, a result it attributes to conflicting learning signals and limited encoder capacity rather than to the structural-supervision idea itself.

Load-bearing premise

The load-bearing premise is that the Gemini-generated switching-point labels and translations are accurate enough that the SPP and BiLTM objectives teach the encoder a genuine structural signal about code-mixing rather than a model of the annotating model's artifacts.

Editorial extensions

If this is right

  • An encoder pretrained with translation and switching-point objectives can be fine-tuned for downstream code-mixed classification at the same inference cost as a standard BERT encoder, since the decoders are detached after pretraining.
  • Switching-point supervision could be applied to other code-mixed language pairs without parallel corpora, as long as word-level language labels or an LLM annotator can supply switch positions.
  • The reported degradation from adding BTSP, TLC, and CMI suggests that future multi-task code-mixed pretraining should treat objectives as competing and tune loss weights and encoder size rather than assuming all auxiliary tasks help.
  • Pretraining on only 10,000 sentences is claimed to be sufficient to produce measurable downstream gains, which would make targeted structural pretraining a cheap alternative to large-scale code-mixed corpus collection.

Reading between the lines

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

  • A direct test of the central claim would replace the LLM-generated switching labels with human-annotated labels on a held-out subset and check whether the HASOC F1 gain persists; if it does not, the SPP gain is tied to the label generator rather than to switching structure.
  • The architecture could transfer to other high-resource code-mixed pairs, but the paper only evaluates Hinglish; a reader should not assume similar gains for pairs with different typology or less Roman-script standardization.
  • Since the attention analysis is qualitative, a quantitative metric—such as the correlation between attention scores and ground-truth switch positions across layers—could turn the claimed switching sensitivity into a falsifiable intrinsic evaluation.
  • The observation that BTSP, TLC, and CMI hurt performance suggests the multi-task objective space is not monotone; tuning per-objective weights or using a larger encoder could flip those signs, an experiment the paper explicitly lists as future work.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 proposes CMLFormer, a Transformer with a shared encoder and two synchronously cross-attending decoders, pre-trained on 10,000 sentences sampled from L3Cube-HingCorpus and augmented with Gemini-generated English/Hindi translations and token-level language/switching-point annotations. Six pre-training objectives are combined (MLM, BTSP, BiLTM, TLC, SPP, CMI). After pre-training, the decoders are detached and the encoder is fine-tuned on HASOC-2021 hate-speech detection. Table 1 reports that the MLM+BiLTM configuration reaches F1=0.431 versus F1=0.249 for a locally pre-trained BERTbase, with F1 degrading as further objectives are added; attention visualizations are used to argue that the model attends to switching points. The paper claims F1 gains of 0.18 and 0.05 from BiLTM and SPP, respectively.

Significance. If the best result were robust, the dual-decoder architecture with code-mixing-specific objectives would be a useful contribution to code-mixed NLP, which currently lacks strong encoder representations for languages like Hinglish. The paper is transparent about its main limitations (small pre-training sample, underparameterized encoder) and ships its tokenizer, code, and data-augmentation prompt, which is a concrete reproducibility asset. The significance is tempered, however, by the absence of error bars, official or published baselines, and a clean ablation isolating the architectural contribution from the objective contribution. The central claim therefore rests on evidence that is internally inconsistent in one case and confounded in another.

major comments (5)
  1. [§6.2, Table 1] The text claims that BiLTM and SPP yield F1 gains of 0.18 and 0.05, respectively, but Table 1 contradicts the ordering. The MLM+BiLTM row reaches F1=0.431; adding SPP to that configuration reduces F1 to 0.295, a drop of 0.136. The 0.046 difference between the MLM+BiLTM+SPP row and the BERTbase row cannot be attributed to SPP because no MLM+SPP condition is reported. Section 9 repeats this unsupported attribution and should be revised or backed by the missing ablation.
  2. [§6.2, Table 1] The claimed BiLTM gain is confounded with the architectural change. Every CMLFormer row uses the dual-decoder architecture with cross-decoder attention, whereas the BERTbase row uses a standard single-stream encoder; no dual-decoder CMLFormer with MLM only is reported. The contribution of BiLTM is therefore not separable from the added decoder parameters and cross-attention mechanism, and an MLM-only CMLFormer ablation is required to support the attribution.
  3. [§3.5.2 and §5] The claim of outperforming "known baselines" is not established. The comparison model is a BERTbase pre-trained locally on the same 10,000-sentence sample (footnote 2), not the official HingBERT released with L3Cube-HingCorpus, and no published HASOC-2021 results from prior shared-task systems or models such as MuRIL are cited. Without those external comparisons, the headline numbers do not support the phrase "outperforms known baselines."
  4. [§6.2 and Table 1] The results are reported without error bars, repeated seeds, or significance testing. Given the small pre-training corpus (10,000 sentences) and the moderate fine-tuning set (5,740 training examples), deltas such as the accuracy differences of 0.008 and 0.002 in rows 2 and 3 are within ordinary optimization noise. The central performance claim needs at least multiple random seeds with standard deviations or a paired significance test.
  5. [§6.3 and Appendix A.1.3] The attention analysis is circular as evidence of switching-point sensitivity. The model was explicitly trained to predict the same Gemini-generated switching labels through the SPP objective, so heightened attention at those positions is an expected consequence of supervised training rather than an independent demonstration of emergent code-mixed structure. In addition, the quality of the LLM-generated translations and token-level labels is only manually spot-checked; no quantitative label-quality metric (e.g., agreement with human annotation) is provided, so the auxiliary objectives may propagate label noise.
minor comments (6)
  1. [Table 1 caption] The caption uses "CLM" where the text and table header use "BiLTM"; this inconsistency should be fixed.
  2. [Appendix A.2] The architecture ablations are described in the future tense ("we will decouple", "we will investigate") rather than as completed experiments; they should either be actually run or explicitly labeled as planned work.
  3. [Equation (2) and Table 3] The CMI weighting factors w_n and w_p and the objective-loss weights α, β, γ, η, ζ, δ are free parameters, but no sensitivity analysis or justification for the chosen values is provided.
  4. [Footnote 1 and Section 8] The dataset size is given as 52.93M sentences in footnote 1 and as 52.9M in Section 8; the inconsistency should be corrected.
  5. [Section 6.3, Figure 3] The attention aggregation procedure (layer, head, and normalization) is not specified in enough detail to reproduce Figure 3; the same extraction should be applied to BERTbase and CMLFormer.
  6. [Appendix A.1.3] The appendix ends with a duplicated period ("switching points.."), and Section 6.2 contains the typo "Bilingual Translation Language Modeling" instead of "Bilingual Language Translation Modeling."

Circularity Check

1 steps flagged · score 6.0 of 10

SPP attention validation is circular; external HASOC benchmark remains independent.

  1. fitted input called prediction [Section 6.3 (Learning Switching Point Dynamics); cf. Section 3.4.5 and Appendix A.1.3]
    "Our results show that CMLFormer far exceeds BERTbase’s capability at identifying switching points in code-mixed inputs. Our switching point focused pre-training objectives enable it to consistently attend to language transitions, giving it significantly higher weights compared to other tokens."

    Section 3.4.5 defines SPP as a supervised token-level binary classification against ground-truth switching labels T, and Appendix A.1.3 says the labels were used to explicitly teach the model to recognize and anticipate language switching points. The Section 6.3 attention analysis then uses those same labels as evidence that CMLFormer can identify switching points, validating its sensitivity to code-mixed structure. Because the encoder was directly optimized on these labels, attention peaks at the annotated switching points are a learned response to the training signal, not an independent confirmation of structural understanding.

full rationale

The paper's principal downstream claim is evaluated on the external HASOC-2021 benchmark against a BERTbase baseline pre-trained on the same sample, so the reported F1 improvements are not circular: the fine-tuning labels are independent of the pre-training objectives. The architecture and multi-task losses are stated with explicit equations and do not presuppose the benchmark result. The one load-bearing circularity is the intrinsic attention evaluation in Section 6.3: the model's ability to 'identify' switching points is presented as validation, but SPP was trained on exactly those Gemini-generated switching labels, and BERTbase was not given SPP supervision, so the attention contrast is a consequence of the training setup rather than an emergent finding. This does not undermine the external benchmark comparison, but it does mean one of the paper's supporting validations reduces to its own supervised input. Separately, the conclusion's attribution of a 0.05 F1 gain to SPP is contradicted by Table 1, where adding SPP to the BiLTM configuration lowers F1 from 0.431 to 0.295; that is a numerical or correctness issue, not a circularity. Overall score reflects partial circularity in the attention analysis while the central HASOC result retains independent content.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the reliability of the LLM-generated augmentation pipeline, the representativeness of the 10k-sentence sample, the correctness of the HASOC labels, and the chosen loss weights. None of these are independently verified in the paper, and the loss weights and CMI factors are set by hand without sensitivity analysis.

free parameters (3)
  • Pre-training loss weights (alpha, beta, gamma, eta, zeta, delta) = 1.0, 1.0, 10.0, 1.0, 10.0, 1.0 (Table 3)
    Chosen by hand without reported sensitivity analysis; the relative weighting of objectives affects the downstream F1 trade-off shown in Table 1.
  • CMI weighting factors w_n and w_p = not stated
    Introduced in Eq. (2) to balance token distribution and transition frequency; values are never reported, so the CMI labels used for the CMIP objective are underspecified.
  • Pre-training sample size = 10,000 sentences
    Chosen due to computational limits (footnote 1); the paper acknowledges this invites overfitting and is a load-bearing constraint on generalization claims.
assumptions (5)
  • ad hoc to paper Gemini 2.0 Flash produces accurate translations and token-level language annotations for Hinglish.
    The entire augmented pre-training corpus and all switching point labels used for SPP and BiLTM are generated by this API (Appendix A.1); no human verification beyond a few hundred samples is described.
  • domain assumption L3Cube-HingCorpus is a representative sample of natural Hinglish.
    The paper pre-trains on 10,000 sentences sampled from this corpus and assumes it reflects real-world code-mixing patterns.
  • domain assumption HASOC-2021 labels are ground truth for hate speech detection.
    Used as the external benchmark; the paper assumes the labels are correct.
  • domain assumption The base-language hypothesis (Hindi provides grammar, English contributes loan words).
    Motivates the dual-decoder split into base and mixing languages; if this decomposition is wrong, the architecture's premise weakens.
  • standard math Standard Transformer architecture and WordPiece tokenization are suitable for code-mixed text.
    Uses Vaswani et al. (2017) and BERT-style tokenization without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CMLFormer: A Dual Decoder Transformer with Switching Point Learning for Code-Mixed Language Modeling." pith.science (2026). https://pith.science/paper/QDT6BMME

@misc{pith2026250512587,
  author       = {Pith},
  title        = {Pith review of: CMLFormer: A Dual Decoder Transformer with Switching Point Learning for Code-Mixed Language Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDT6BMME}},
  note         = {Machine review of arXiv:2505.12587}
}
read the original abstract

Code-mixed languages, characterized by frequent within-sentence language transitions, present structural challenges that standard language models fail to address. In this work, we propose CMLFormer, an enhanced multi-layer dual-decoder Transformer with a shared encoder and synchronized decoder cross-attention, designed to model the linguistic and semantic dynamics of code-mixed text. CMLFormer is pre-trained on an augmented Hinglish corpus with switching point and translation annotations with multiple new objectives specifically aimed at capturing switching behavior, cross-lingual structure, and code-mixing complexity. Our experiments show that CMLFormer improves F1 score, precision, and accuracy over other approaches on the HASOC-2021 benchmark under select pre-training setups. Attention analyses further show that it can identify and attend to switching points, validating its sensitivity to code-mixed structure. These results demonstrate the effectiveness of CMLFormer's architecture and multi-task pre-training strategy for modeling code-mixed languages.

Figures

Figures reproduced from arXiv: 2505.12587 by the authors.

Figure 1
Figure 1. Switching Points in Code-Mixed Text alter structural dynamics. Code-mixing introduces switch￾ing points where the language transitions mid-sentence, significantly affecting token order and grammatical structure. In the example shown, words like tomorrow (originally after attend in English) now appear earlier in the Hindi-English mixed variant, changing the sen￾tence’s grammatical flow, which multilingual models stru… view at source ↗
Figure 2
Figure 2. The architecture of our proposed approach CMLFormer: The outputs from each encoder-decoder attention sub-layer (arrows shown in green and red) are passed as input to the decoder-decoder cross-attention sub-layer. The decoders exhibit full synchronous cou￾pling since each requires the hidden states from the other to compute its own hidden states. After pre-training, the encoder is extracted and fine-tuned on downstre… view at source ↗
Figure 3
Figure 3. Average Attention Score per Token by CMLFormer and BERTbase. We see that CMLFormer consistently identifies and attends to language transitions around switching points, and is agnostic to the nature and number of transitions; BERT fails to identify any transitions and attends to trivial tokens. 9 Conclusion In this work, we introduce CMLFormer, an en￾hanced dual–decoder Transformer that explicitly targets the linguis… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Loss curve for the total pre-training loss ag￾gregated across all active objectives across different variants of decoder cross-attention. The lower curve for synchronous cross-attention reflects the strength of interaction learned between base and mixing language repre…
Figure 5
Figure 5. Figure 5: Additional attention maps produced by CMLFormer on code-mixed inputs. Attention consistently peaks near switching points [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 12 canonical work pages

  1. [1]

    Mohsin Ali, Kandukuri Sai Teja, Neeharika Gupta, Parth Patwa, Anubhab Chatterjee, Vinija Jain, Aman Chadha, and Amitava Das. 2023. http://arxiv.org/abs/2309.05270 Conflator: Incorporating switching point based rotatory positional encodings for code-mixed language modeling

    work page Pith review arXiv 2023

  2. [2]

    Mohsin Ali, Kandukuri Sai Teja, Sumanth Manduru, Parth Patwa, and Amitava Das. 2021. http://arxiv.org/abs/2111.06599 Pesto: Switching point based dynamic and relative positional encoding for code-mixed languages

    work page Pith review arXiv 2021

  3. [3]

    Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Gemini Team, Jiahui Yu, and Radu Soricut et. al. 2024. http://arxiv.org/abs/2312.11805 Gemini: A family of highly capable multimodal models

    arXiv 2024

  4. [4]

    o rn Gamb \

    Amitava Das and Bj \"o rn Gamb \"a ck. 2014. Identifying languages at the word level in code-mixed indian social media text. In Proceedings of the 11th International Conference on Natural Language Processing (ICON-2014), pages 169--178

    work page 2014

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. http://arxiv.org/abs/1810.04805 Bert: Pre-training of deep bidirectional transformers for language understanding

    arXiv 2019

  6. [6]

    Khapra, Anoop Kunchukuttan, and Pratyush Kumar

    Sumanth Doddapaneni, Rahul Aralikatte, Gowtham Ramesh, Shreya Goyal, Mitesh M. Khapra, Anoop Kunchukuttan, and Pratyush Kumar. 2023. http://arxiv.org/abs/2212.05409 Towards leaving no indic language behind: Building monolingual corpora, benchmark and models for indic languages

    arXiv 2023

  7. [7]

    Guerreiro, António Loison, Duarte M

    Manuel Faysse, Patrick Fernandes, Nuno M. Guerreiro, António Loison, Duarte M. Alves, Caio Corro, Nicolas Boizard, João Alves, Ricardo Rei, Pedro H. Martins, Antoni Bigata Casademunt, François Yvon, André F. T. Martins, Gautier Viaud, Céline Hudelot, and Pierre Colombo. 2024. http://arxiv.org/abs/2402.00786 Croissantllm: A truly bilingual french-english l...

    arXiv 2024

  8. [8]

    Shruti Jagdale, Omkar Khade, Gauri Takalikar, Mihir Inamdar, and Raviraj Joshi. 2024. http://arxiv.org/abs/2411.18577 On importance of code-mixed embeddings for hate speech identification

    work page Pith review arXiv 2024

Show all 24 references
  1. [9]

    Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja Nagipogu, Shachi Dave, Shruti Gupta, Subhash Chandra Bose Gali, Vish Subramanian, and Partha Talukdar. 2021. http://arxiv.org/abs/2103.1073...

    2021 arXiv

  2. [10]

    Guillaume Lample and Alexis Conneau. 2019. http://arxiv.org/abs/1901.07291 Cross-lingual language model pretraining

    2019 arXiv

  3. [11]

    Shuyue Stella Li and Kenton Murray. 2022. http://arxiv.org/abs/2211.07628 Language agnostic code-mixing data augmentation by predicting linguistic patterns

    2022 arXiv

  4. [12]

    Thomas Mandl, Sandip Modha, Gautam Kishore Shahi, Hiren Madhu, Shrey Satapara, Prasenjit Majumder, Johannes Schaefer, Tharindu Ranasinghe, Marcos Zampieri, Durgesh Nandini, and Amit Kumar Jaiswal. 2021. http://arxiv.org/abs/2112.09301 Overview of the hasoc subtrack at fire 202...

    2021 arXiv

  5. [13]

    Debajyoti Mazumder, Aakash Kumar, and Jasabanta Patro. 2024. http://arxiv.org/abs/2412.12761 Revealing the impact of synthetic native samples and multi-tasking strategies in hindi-english code-mixed humour and sarcasm detection

    2024 arXiv

  6. [14]

    Ravindra Nayak and Raviraj Joshi. 2022. https://aclanthology.org/2022.wildre-1.2 L 3 C ube- H ing C orpus and H ing BERT : A code mixed H indi- E nglish dataset and BERT language models . In Proceedings of the WILDRE-6 Workshop within the 13th Language Resources and Evaluation...

    2022

  7. [15]

    Aryan Patil, Varad Patwardhan, Abhishek Phaltankar, Gauri Takawane, and Raviraj Joshi. 2023. https://doi.org/10.1109/i2ct57861.2023.10126273 Comparative study of pre-trained bert models for code-mixed hindi-english data . In 2023 IEEE 8th International Conference for Convergen...

    2023

  8. [16]

    Sebastin Santy, Anirudh Srinivasan, and Monojit Choudhury. 2021. https://aclanthology.org/2021.adaptnlp-1.12 BERT ologi C o M ix: How does code-mixing interact with multilingual BERT ? In Proceedings of the Second Workshop on Domain Adaptation for NLP, pages 111--121, Kyiv, Uk...

    2021

  9. [17]

    Shad Akhtar

    Ayan Sengupta, Sourabh Kumar Bhattacharjee, Tanmoy Chakraborty, and Md. Shad Akhtar. 2021. https://doi.org/10.18653/v1/2021.findings-acl.407 HIT - a hierarchically fused deep attention network for robust code-mixed language representation . In Findings of the Association for C...

    2021 doi

  10. [18]

    Varad Srivastava. 2025. https://aclanthology.org/2025.chipsal-1.5/ D wesh V aani: An LLM for detecting religious hate speech in code-mixed H indi- E nglish . In Proceedings of the First Workshop on Challenges in Processing South Asian Languages (CHiPSAL 2025), pages 46--60, Ab...

    2025

  11. [19]

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2023. http://arxiv.org/abs/2104.09864 Roformer: Enhanced transformer with rotary position embedding

    2023 arXiv

  12. [20]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Neural Informa...

    2017

  13. [21]

    Dilip Venkatesh, Pasunti Prasanjith, and Yashvardhan Sharma. 2024. https://doi.org/10.18653/v1/2024.semeval-1.115 BITS pilani at S em E val-2024 task 10: Fine-tuning BERT and llama 2 for emotion recognition in conversation . In Proceedings of the 18th International Workshop on...

    2024 doi

  14. [22]

    Ruochen Zhang, Samuel Cahyawijaya, Jan Christian Blaise Cruz, Genta Indra Winata, and Alham Fikri Aji. 2023. http://arxiv.org/abs/2305.14235 Multilingual large language models are not (yet) code-switchers

    2023 arXiv

  15. [23]

    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...

  16. [24]

    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 15, 2026 · model on record in the stance chip above.