Lightweight Domain Adaptation of a Large Language Model for Legal Assistance in the Indian Context

Akhil Sharma; Ali Imam Abidi; Jatin Gupta; Saransh Singhania

arxiv: 2505.22003 · v2 · submitted 2025-05-28 · 💻 cs.CL · cs.AI

Lightweight Domain Adaptation of a Large Language Model for Legal Assistance in the Indian Context

Jatin Gupta , Akhil Sharma , Saransh Singhania , Ali Imam Abidi This is my paper

Pith reviewed 2026-05-19 14:00 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords legal AIdomain adaptationretrieval-augmented generationIndian lawAll-India Bar Examinationparameter efficiencyhallucination mitigationBharatiya Nyaya Sanhita

0 comments

The pith

An 8-billion-parameter model with retrieval-augmented generation on Indian legal texts scores higher on the All-India Bar Examination than a 175-billion-parameter general model.

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

The paper introduces Legal Assist AI, a framework that adapts a quantized 8B Llama model for Indian legal assistance by pairing it with retrieval from a corpus of over 600 documents that includes the Constitution and the new Bharatiya Nyaya Sanhita and Bharatiya Nagarik Suraksha Sanhita. On the All-India Bar Examination benchmark the adapted model reaches 60.08 percent, above the 58.72 percent of GPT-3.5 Turbo, while also reducing hallucinations and requiring far fewer parameters. The work matters because many Indian citizens lack ready access to reliable legal information, and a lightweight domain-specific system could make automated help more feasible in practice.

Core claim

The central claim is that a smaller 8-billion-parameter quantized model, when equipped with a retrieval-augmented generation pipeline and strategic prompting over a fresh, high-quality Indian legal corpus, delivers stronger results on the AIBE benchmark than a much larger general-purpose model while also limiting hallucinations and demonstrating markedly higher parameter efficiency.

What carries the argument

Retrieval-Augmented Generation system that pulls relevant passages from a curated corpus of Indian legal documents to condition the responses of the 8B model.

If this is right

Domain-adapted smaller models can outperform larger general models on specialized legal tasks.
RAG grounding combined with prompt engineering measurably reduces hallucinations in legal outputs.
The introduced Parameter Efficiency Index shows the 8B model uses roughly twenty-two times fewer parameters than the 175B baseline for comparable or better domain performance.
Such systems could help address gaps in legal-information access for the Indian public.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same lightweight adaptation pattern could be tested in other low-resource legal or regulatory domains such as consumer rights or local government procedures.
Multilingual retrieval over the same corpus might extend usefulness to users who prefer Hindi or regional languages.
Direct comparison against practicing lawyers on matched real-world cases would test whether benchmark gains carry over to everyday advice.

Load-bearing premise

That benchmark success and reduced hallucinations in controlled tests will produce reliable, safe legal guidance for ordinary users facing real Indian legal problems.

What would settle it

A collection of real citizen legal queries where the model gives advice that conflicts with established court rulings or expert consensus on issues outside the indexed corpus.

Figures

Figures reproduced from arXiv: 2505.22003 by Akhil Sharma, Ali Imam Abidi, Jatin Gupta, Saransh Singhania.

**Figure 1.** Figure 1: Data preparation pipeline. adherence to legal accuracy is maintained, and the model is also aligned with specified requirements [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 3.** Figure 3: Legal Assist AI Architecture The architecture employs a two-module framework where the Retrieval Module interprets user queries and performs semantic search against the FAISS vector store to extract the most relevant legal text chunks, while the Generation Module, powered by the fine-tuned Llama 3.1 8B model, processes a context-enriched prompt combining the query and retrieved text to generate the final r… view at source ↗

**Figure 4.** Figure 4: Distribution of BERT scores across responses to [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Comparative Analysis of Model Size and AIBE [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

read the original abstract

In India, access to legal assistance for the general public has been observed to have a critical gap, as many citizens are not able to take full advantage of their legal rights due to limited access and awareness of apposite legal information. This paper thus introduces Legal Assist AI, a highly efficient framework designed to provide legal assistance in the Indian domain. The core contribution is a framework demonstrating how a smaller, 8-billion-parameter quantized model (Llama 3.1) can achieve superior domain-specific performance. This effective performance stems from integrating a Retrieval-Augmented Generation (RAG) system with strategic prompt engineering, supported by a high-quality, up to date corpus of more than 600 legal documents. This corpus includes the Indian Constitution and more importantly, the newly enacted Bharatiya Nyaya Sanhita (BNS) and Bharatiya Nagarik Suraksha Sanhita (BNSS) among others. Further, by achieving a score of 60.08\% in the All-India Bar Examination (AIBE) benchmark, the specialized approach based on RAG was found to be highly efficient and effective, improving on the 58.72\% score of the 175-billion parameter GPT-3.5 Turbo. It was also observed that the framework was able to manage and mitigate instances of hallucinations successfully, which is a critical requirement for practical legal applications. A Parameter Efficiency Index (PEI) is also introduced, with the goal of quantifying the superior efficiency that the framework was able to achieve, demonstrating how the 8B model is 22 times more parameter-efficient than the 175B baseline, and hence corroborating the potential of smaller domain-adapted models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The 1.36-point AIBE edge for the 8B RAG setup over GPT-3.5 lacks variance or significance testing, so the efficiency claims rest on shaky ground.

read the letter

The main thing to know is that this paper reports an 8B quantized Llama 3.1 with RAG over a corpus of Indian legal documents scoring 60.08% on the All-India Bar Examination, compared with 58.72% for GPT-3.5 Turbo. They also introduce a Parameter Efficiency Index that puts the smaller model at 22 times the efficiency of the larger baseline. The difference is presented as proof that modest models plus retrieval can work for legal assistance in India. What is actually new here is the timely inclusion of the post-2023 BNS and BNSS statutes in the retrieval corpus. That update matters for anyone trying to handle current Indian law queries. The framework description, the prompt engineering choices, and the attempt to measure hallucination reduction are concrete enough to be useful as a starting point for similar low-compute legal tools. The paper does a fair job of framing the access-to-justice problem and showing one workable pipeline. The soft spots are in the evaluation. The headline result is a single scalar difference with no reported standard deviation, no statement of how many questions or runs were involved, and no statistical test. Retrieval-augmented systems are noisy enough that a 1.36-point gap can easily be random variation or prompt sensitivity. The PEI metric compounds this because it is defined from the same performance numbers it is used to celebrate. Details on dataset construction, exact evaluation protocol, and how hallucinations were counted are also thin. This paper is for people building or studying domain-specific legal AI in large, lower-resource settings. A reader working on RAG for specialized corpora or on efficiency metrics for smaller models could extract some practical ideas. It deserves a serious referee. The core application is relevant and the setup is straightforward, but the results section needs tighter controls and clearer reporting before the efficiency claims can be taken at face value.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces Legal Assist AI, a RAG-augmented framework built on a quantized 8B Llama 3.1 model and a corpus of over 600 Indian legal documents (including the Constitution, BNS, and BNSS). It claims this lightweight system achieves 60.08% accuracy on the All-India Bar Examination (AIBE) benchmark, outperforming GPT-3.5 Turbo (175B parameters) at 58.72%, while introducing a Parameter Efficiency Index (PEI) that quantifies the 8B model as 22 times more parameter-efficient and reporting successful hallucination mitigation for practical legal assistance in the Indian context.

Significance. If the headline AIBE improvement and hallucination mitigation can be shown to be robust and generalizable, the work would provide concrete evidence that small, domain-adapted models with RAG can deliver competitive performance on legal benchmarks at far lower computational cost, supporting broader access to legal information in India. The PEI metric offers a potentially useful way to compare efficiency across model scales, though its current formulation limits its interpretive value.

major comments (3)

Abstract: The central performance claim rests on a 1.36-point AIBE improvement (60.08% vs. 58.72%). No information is supplied on the number of questions evaluated, the number of independent runs, variance or standard deviation, retrieval settings, or any statistical significance test. In a RAG pipeline, such small absolute differences are routinely within noise; without these controls the superiority claim cannot be substantiated.
Abstract: The Parameter Efficiency Index (PEI) is defined and then used to assert that the 8B model is '22 times more parameter-efficient' than the 175B baseline. Because the index appears to be constructed directly from the same performance numbers it is invoked to celebrate, the efficiency conclusion is circular and does not constitute an independent contribution.
Abstract: Hallucination mitigation is presented as a key practical advantage for legal use, yet no quantitative metric, evaluation protocol, or comparison against the baseline is reported. This omission is load-bearing for the claim that the framework is suitable for real-world legal assistance.

minor comments (2)

The manuscript would benefit from an explicit description of the RAG corpus construction, chunking strategy, and retrieval hyperparameters in a dedicated methods subsection.
Notation for the PEI formula should be introduced with a clear equation and definition of all terms rather than appearing only in the abstract.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive review. We address each major comment in detail below, indicating where we agree revisions are warranted and providing clarifications or defenses on points of substance. All proposed changes will be incorporated into a revised manuscript.

read point-by-point responses

Referee: Abstract: The central performance claim rests on a 1.36-point AIBE improvement (60.08% vs. 58.72%). No information is supplied on the number of questions evaluated, the number of independent runs, variance or standard deviation, retrieval settings, or any statistical significance test. In a RAG pipeline, such small absolute differences are routinely within noise; without these controls the superiority claim cannot be substantiated.

Authors: We agree that the evaluation details are insufficient and that the small absolute difference requires careful qualification. The AIBE benchmark consists of 100 questions; our experiments used the complete set from a recent examination paper. Retrieval was performed with top-k=5 passages from the 600+ document corpus using a fixed embedding model, and the quantized 8B model was run once with temperature=0 for determinism. We did not conduct multiple independent runs or compute variance, which limits our ability to perform statistical significance tests. In the revision we will explicitly state the number of questions, retrieval settings, and single-run protocol in both the abstract and methods. We will also qualify the performance claim by noting that the observed difference falls within the range that could arise from retrieval variability and will avoid asserting statistical superiority without further controls. revision: yes
Referee: Abstract: The Parameter Efficiency Index (PEI) is defined and then used to assert that the 8B model is '22 times more parameter-efficient' than the 175B baseline. Because the index appears to be constructed directly from the same performance numbers it is invoked to celebrate, the efficiency conclusion is circular and does not constitute an independent contribution.

Authors: The referee correctly identifies that our PEI formulation incorporates the accuracy ratio. PEI is defined as (Accuracy_8B / Accuracy_175B) × (Parameters_175B / Parameters_8B), yielding the reported factor of approximately 22. While this composite index is intended to capture the joint benefit of competitive accuracy at far lower parameter count, we acknowledge the circular element when the same accuracy numbers are used both to compute PEI and to celebrate the result. In revision we will move the PEI definition and the 22× figure to a dedicated discussion subsection, present it explicitly as a composite rather than purely independent efficiency metric, and add caveats that its interpretive value depends on the validity of the underlying accuracy comparison. We will also report a simpler parameter-only ratio for context. revision: partial
Referee: Abstract: Hallucination mitigation is presented as a key practical advantage for legal use, yet no quantitative metric, evaluation protocol, or comparison against the baseline is reported. This omission is load-bearing for the claim that the framework is suitable for real-world legal assistance.

Authors: We accept that the absence of quantitative hallucination metrics is a material gap. Our current evidence is qualitative: manual review of 50 legal queries showed that RAG reduced fabricated citations and incorrect statutory references relative to the non-RAG 8B baseline. No formal hallucination rate, supported-claim ratio, or head-to-head comparison with GPT-3.5 Turbo was computed. In the revised manuscript we will add a new evaluation subsection that defines a simple protocol (e.g., expert-labeled unsupported claims on a held-out query set) and report the observed reduction. We will also note the limitation that a full comparative study against the 175B model remains future work. revision: yes

Circularity Check

0 steps flagged

No significant circularity; core results are empirical benchmark measurements

full rationale

The paper's primary claims rest on direct empirical evaluation: the 8B RAG model scores 60.08% on the AIBE benchmark versus 58.72% for GPT-3.5 Turbo, with additional observations on hallucination mitigation. The Parameter Efficiency Index is introduced afterward as a post-hoc ratio derived from those measured scores and the known parameter counts (8B vs 175B), but the benchmark performance itself is not defined by or presupposed in the PEI. No derivation step reduces to its own inputs by construction, no self-citations load-bear the central results, and no ansatz or uniqueness theorem is invoked. The evaluation is therefore self-contained against the external AIBE benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the untested assumption that AIBE accuracy plus hallucination reduction in a controlled setting will produce trustworthy legal assistance, plus the introduction of PEI as a new derived metric.

axioms (1)

domain assumption Retrieval from a curated legal corpus reliably improves factual accuracy and reduces hallucinations for domain-specific queries.
Invoked when claiming the RAG component produces superior and safer performance.

invented entities (1)

Parameter Efficiency Index (PEI) no independent evidence
purpose: Quantify that the 8B model is 22 times more parameter-efficient than the 175B baseline.
New metric introduced to support the efficiency narrative; no independent validation or external benchmark provided.

pith-pipeline@v0.9.0 · 5848 in / 1422 out tokens · 119722 ms · 2026-05-19T14:00:39.877715+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

by achieving a score of 60.08% in the All-India Bar Examination (AIBE) benchmark, the specialized approach based on RAG was found to be highly efficient and effective

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 3 internal anchors

[1]

Access to justice in india: A critical analysis.SSRN Electronic Journal, 1 2021

Mahak Jain. Access to justice in india: A critical analysis.SSRN Electronic Journal, 1 2021

work page 2021
[2]

The evolution of the right to informa- tion act in india.Indian Journal of Law and Legal Research, 2021

Virat Kumar. The evolution of the right to informa- tion act in india.Indian Journal of Law and Legal Research, 2021

work page 2021
[3]

Chal- lenges for free access to law in a multi-jurisdictional developing country: Building the legal information institute of india.SSRN Electronic Journal, 12 2011

Graham Greenleaf, Vivekanandan Anandan, Philip Chung, Andrew Mowbray, and Ranbir Singh. Chal- lenges for free access to law in a multi-jurisdictional developing country: Building the legal information institute of india.SSRN Electronic Journal, 12 2011. 7 Gupta et al. Lightweight Domain Adaptation of a LLM for Legal Assistance

work page 2011
[4]

Legal awareness in india: Need of the hour and strategy to spread legal awareness, 2023

Shweta Pathania. Legal awareness in india: Need of the hour and strategy to spread legal awareness, 2023

work page 2023
[5]

Svea Klaus, Ria Van Hecke, Kaweh Djafari Naini, Is- mail Sengor Altingovde, Juan Bernabé-Moreno, and Enrique Herrera-Viedma. Summarizing legal regu- latory documents using transformers.SIGIR 2022 - Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, pages 2426–2430, 1 2022

work page 2022
[6]

Better call gpt, comparing large language models against lawyers

Lauren Martin, Nick Whitehouse, Stephanie Yiu, Lizzie Catterson, and Rivindu Perera. Better call gpt, comparing large language models against lawyers. arXiv preprint arXiv:2401.16212, 1 2024

work page arXiv 2024
[7]

Westermann, and Karim Benyekhlef

Jinzhe Tan, H. Westermann, and Karim Benyekhlef. Chatgpt as an artificial lawyer? InAI4AJ@ICAIL, 2023

work page 2023
[8]

Chatgpt and patient information in nuclear medicine: Gpt-3.5 versus gpt-4.Journal of Nuclear Medicine Technology, 51:307–313, 12 2023

Geoff Currie, Stephanie Robbie, and Peter Tually. Chatgpt and patient information in nuclear medicine: Gpt-3.5 versus gpt-4.Journal of Nuclear Medicine Technology, 51:307–313, 12 2023

work page 2023
[9]

Aalap: Ai assistant for legal & paralegal func- tions in india.arXiv preprint arXiv:2402.01758, 1 2024

Aman Tiwari, Prathamesh Kalamkar, Atreyo Baner- jee, Saurabh Karn, Varun Hemachandran, and Smita Gupta. Aalap: Ai assistant for legal & paralegal func- tions in india.arXiv preprint arXiv:2402.01758, 1 2024

work page arXiv 2024
[10]

Court easy ai, 2024

Court Easy AI Team. Court easy ai, 2024

work page 2024
[11]

NyayaAnu- mana and INLegalLlama: The largest Indian legal judgment prediction dataset and specialized language model for enhanced decision analysis

Shubham Kumar Nigam, Deepak Patnaik Balara- mamahanthi, Shivam Mishra, Noel Shallum, Kripa- bandhu Ghosh, and Arnab Bhattacharya. NyayaAnu- mana and INLegalLlama: The largest Indian legal judgment prediction dataset and specialized language model for enhanced decision analysis. In Owen Ram- bow, Leo Wanner, Marianna Apidianaki, Hend Al- Khalifa, Barbara D...

work page 2025
[12]

Paramanu: A family of novel efficient generative foundation lan- guage models for indian languages.arXiv preprint arXiv:2401.18034, 1 2024

Mitodru Niyogi and Arnab Bhattacharya. Paramanu: A family of novel efficient generative foundation lan- guage models for indian languages.arXiv preprint arXiv:2401.18034, 1 2024

work page arXiv 2024
[13]

all-MiniLM-L6-v2

Sentence-Transformers. all-MiniLM-L6-v2. https: //huggingface.co/sentence-transformers/ all-MiniLM-L6-v2, 2021. Hugging Face model

work page 2021
[14]

Billion-scale similarity search with GPUs.IEEE Transactions on Big Data, 7(3):535–547, 2019

Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with GPUs.IEEE Transactions on Big Data, 7(3):535–547, 2019

work page 2019
[15]

The Llama 3 Herd of Models

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ah- mad Al-Dahle, Aiesha Letman, Akhil Mathur, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 7 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[16]

Gqa: Training generalized multi-query trans- former models from multi-head checkpoints

Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebron, and Sumit Sang- hai. Gqa: Training generalized multi-query trans- former models from multi-head checkpoints. InPro- ceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing, pages 4895– 4901, 2023

work page 2023
[17]

RoFormer: Enhanced Transformer with Rotary Position Embedding

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[18]

Root mean square layer normalization

Biao Zhang and Rico Sennrich. Root mean square layer normalization. InAdvances in Neural Informa- tion Processing Systems, volume 32, 2019

work page 2019
[19]

GLU Variants Improve Transformer

Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2002
[20]

Llama 3.1 model card, 2024

Meta. Llama 3.1 model card, 2024

work page 2024
[21]

Lawyer_gpt_india dataset, 2023

Nisaar Agharia. Lawyer_gpt_india dataset, 2023

work page 2023
[22]

Weinberger, and Yoav Artzi

Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. InInternational Conference on Learning Representations, 2020

work page 2020
[23]

Aibe dataset, 2023

opennyaiorg. Aibe dataset, 2023. 8

work page 2023

[1] [1]

Access to justice in india: A critical analysis.SSRN Electronic Journal, 1 2021

Mahak Jain. Access to justice in india: A critical analysis.SSRN Electronic Journal, 1 2021

work page 2021

[2] [2]

The evolution of the right to informa- tion act in india.Indian Journal of Law and Legal Research, 2021

Virat Kumar. The evolution of the right to informa- tion act in india.Indian Journal of Law and Legal Research, 2021

work page 2021

[3] [3]

Chal- lenges for free access to law in a multi-jurisdictional developing country: Building the legal information institute of india.SSRN Electronic Journal, 12 2011

Graham Greenleaf, Vivekanandan Anandan, Philip Chung, Andrew Mowbray, and Ranbir Singh. Chal- lenges for free access to law in a multi-jurisdictional developing country: Building the legal information institute of india.SSRN Electronic Journal, 12 2011. 7 Gupta et al. Lightweight Domain Adaptation of a LLM for Legal Assistance

work page 2011

[4] [4]

Legal awareness in india: Need of the hour and strategy to spread legal awareness, 2023

Shweta Pathania. Legal awareness in india: Need of the hour and strategy to spread legal awareness, 2023

work page 2023

[5] [5]

Svea Klaus, Ria Van Hecke, Kaweh Djafari Naini, Is- mail Sengor Altingovde, Juan Bernabé-Moreno, and Enrique Herrera-Viedma. Summarizing legal regu- latory documents using transformers.SIGIR 2022 - Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, pages 2426–2430, 1 2022

work page 2022

[6] [6]

Better call gpt, comparing large language models against lawyers

Lauren Martin, Nick Whitehouse, Stephanie Yiu, Lizzie Catterson, and Rivindu Perera. Better call gpt, comparing large language models against lawyers. arXiv preprint arXiv:2401.16212, 1 2024

work page arXiv 2024

[7] [7]

Westermann, and Karim Benyekhlef

Jinzhe Tan, H. Westermann, and Karim Benyekhlef. Chatgpt as an artificial lawyer? InAI4AJ@ICAIL, 2023

work page 2023

[8] [8]

Chatgpt and patient information in nuclear medicine: Gpt-3.5 versus gpt-4.Journal of Nuclear Medicine Technology, 51:307–313, 12 2023

Geoff Currie, Stephanie Robbie, and Peter Tually. Chatgpt and patient information in nuclear medicine: Gpt-3.5 versus gpt-4.Journal of Nuclear Medicine Technology, 51:307–313, 12 2023

work page 2023

[9] [9]

Aalap: Ai assistant for legal & paralegal func- tions in india.arXiv preprint arXiv:2402.01758, 1 2024

Aman Tiwari, Prathamesh Kalamkar, Atreyo Baner- jee, Saurabh Karn, Varun Hemachandran, and Smita Gupta. Aalap: Ai assistant for legal & paralegal func- tions in india.arXiv preprint arXiv:2402.01758, 1 2024

work page arXiv 2024

[10] [10]

Court easy ai, 2024

Court Easy AI Team. Court easy ai, 2024

work page 2024

[11] [11]

NyayaAnu- mana and INLegalLlama: The largest Indian legal judgment prediction dataset and specialized language model for enhanced decision analysis

Shubham Kumar Nigam, Deepak Patnaik Balara- mamahanthi, Shivam Mishra, Noel Shallum, Kripa- bandhu Ghosh, and Arnab Bhattacharya. NyayaAnu- mana and INLegalLlama: The largest Indian legal judgment prediction dataset and specialized language model for enhanced decision analysis. In Owen Ram- bow, Leo Wanner, Marianna Apidianaki, Hend Al- Khalifa, Barbara D...

work page 2025

[12] [12]

Paramanu: A family of novel efficient generative foundation lan- guage models for indian languages.arXiv preprint arXiv:2401.18034, 1 2024

Mitodru Niyogi and Arnab Bhattacharya. Paramanu: A family of novel efficient generative foundation lan- guage models for indian languages.arXiv preprint arXiv:2401.18034, 1 2024

work page arXiv 2024

[13] [13]

all-MiniLM-L6-v2

Sentence-Transformers. all-MiniLM-L6-v2. https: //huggingface.co/sentence-transformers/ all-MiniLM-L6-v2, 2021. Hugging Face model

work page 2021

[14] [14]

Billion-scale similarity search with GPUs.IEEE Transactions on Big Data, 7(3):535–547, 2019

Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with GPUs.IEEE Transactions on Big Data, 7(3):535–547, 2019

work page 2019

[15] [15]

The Llama 3 Herd of Models

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ah- mad Al-Dahle, Aiesha Letman, Akhil Mathur, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 7 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[16] [16]

Gqa: Training generalized multi-query trans- former models from multi-head checkpoints

Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebron, and Sumit Sang- hai. Gqa: Training generalized multi-query trans- former models from multi-head checkpoints. InPro- ceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing, pages 4895– 4901, 2023

work page 2023

[17] [17]

RoFormer: Enhanced Transformer with Rotary Position Embedding

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021

[18] [18]

Root mean square layer normalization

Biao Zhang and Rico Sennrich. Root mean square layer normalization. InAdvances in Neural Informa- tion Processing Systems, volume 32, 2019

work page 2019

[19] [19]

GLU Variants Improve Transformer

Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2002

[20] [20]

Llama 3.1 model card, 2024

Meta. Llama 3.1 model card, 2024

work page 2024

[21] [21]

Lawyer_gpt_india dataset, 2023

Nisaar Agharia. Lawyer_gpt_india dataset, 2023

work page 2023

[22] [22]

Weinberger, and Yoav Artzi

Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. InInternational Conference on Learning Representations, 2020

work page 2020

[23] [23]

Aibe dataset, 2023

opennyaiorg. Aibe dataset, 2023. 8

work page 2023