Multi-Intent Recognition in Dialogue Understanding: A Comparison Between Smaller Open-Source LLMs

Adnan Ahmad; Philine Kowol; Sebastian M\"oller; Stefan Hillmann

REVIEW 4 major objections 5 minor 27 references

Multi-Intent Recognition in Dialogue Understanding: A Comparison Between Smaller Open-Source LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Mistral-7B-v0.1 beats Llama2-7B-hf and Yi-6B at few-shot multi-intent classification on MultiWOZ 2.1, but a supervised BERT remains much better.

desk verdict Useful but under-specified empirical comparison: the Mistral-over-Llama/Yi ranking and BERT-supervised superiority are plausible, but the omitted prompt and strict parser make the ranking hard to audit. read the letter →

arxiv 2509.10010 v1 pith:TEVQFOXG submitted 2025-09-12 cs.CL cs.HC

Adnan Ahmad , Philine Kowol , Stefan Hillmann , Sebastian Möller This is my paper · ORCID

classification cs.CLcs.HC

keywords multi-labelintentclassificationfew-shotlearningopen-sourceLLMsMistral-7BLlama2Yi-6BMultiWOZ2.1dialogueunderstanding

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 tests three small open-source LLMs (Llama2-7B-hf, Mistral-7B-v0.1, Yi-6B) on multi-label intent classification for dialogue, using a few-shot prompt with 20 examples from MultiWOZ 2.1. It finds that Mistral-7B-v0.1 has the best F1 for 11 of 14 intent classes and the highest weighted average F1 of 0.50. It also benchmarks a supervised BERT model that reaches 0.92 micro F1, far above the LLMs. The work offers a practical framework for evaluating small open-source models on intent detection tasks in resource-limited settings.

What carries the argument

The few-shot prompting protocol: a system prompt, instruction, list of 14 intent classes with descriptions, and 20 JSON-format examples chosen from the validation set, given to each model at temperature 0. Outputs are parsed into intent sets, with unparsable or hallucinated outputs treated as empty. This protocol is the mechanism that determines each model's predictions, and the comparison is carried out with standard multi-label metrics (micro/macro/weighted F1, Hamming loss, Jaccard). For the supervised baseline, the paper uses BertForSequenceClassification with binary cross-entropy loss and a threshold tuned on validation.

What would settle it

Re-run the same evaluation with a different prompt template (e.g., different instruction wording, different example selection, or chain-of-thought phrasing) on the same MultiWOZ 2.1 test set; if a model other than Mistral achieves a higher weighted F1, the paper's main ranking claim fails. Likewise, changing the parsing rule to count out-of-vocabulary intents as errors rather than empty outputs could shift the scores.

Watch

Extended reading notes

Core claim

The paper's central finding is that among the three tested 7B-class open-source generative models, Mistral-7B-v0.1 is the most accurate for few-shot, multi-label intent classification of user utterances in the MultiWOZ 2.1 benchmark. It achieves a weighted F1 of 0.50, outperforming the other models on 11 out of 14 intent classes, and it has the lowest Hamming loss and highest Jaccard similarity of the three. The authors also show that a supervised BERT classifier trained on the full training set (with or without prior system-utterance context) substantially outperforms the best few-shot LLM, reaching a micro F1 of 0.92 versus 0.45 for Mistral.

Load-bearing premise

The hand-tuned prompt and the 20 examples, selected through iterative experiments on the validation set, are assumed to be fair and near-optimal for all three models, so that the observed ranking reflects model capability rather than prompt sensitivity.

Editorial extensions

If this is right

Mistral-7B-v0.1 can serve as a ready-to-use few-shot intent classifier for task-oriented dialogue systems without fine-tuning, provided that the user's task resembles the MultiWOZ distribution and prompt format.
For developers with access to training data, a supervised BERT classifier (110M parameters) offers much higher accuracy and far faster evaluation, making it the preferred option in standard settings.
The comparison framework—fixed prompt, 20 examples, temperature 0, deterministic parsing—can be reused to evaluate other small open-source LLMs on multi-label intent classification.
The gap between supervised and few-shot performance suggests that few-shot generative models are currently not sufficient for high-accuracy NLU pipelines, but they are a viable zero-training-data fallback.
Model choice matters: Yi-6B, despite being competitive on generic leaderboards, performs the worst on this task, indicating task-specific evaluation is necessary.

Reading between the lines

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

The ranking of the three models may be sensitive to prompt wording and the specific 20 examples chosen; if a different hand-tuned prompt or example set were used, another model might win. This is an implicit limitation the paper acknowledges.
The parser's treatment of unparsable outputs as empty labels likely penalizes models that express intents in non-constrained formats; a more lenient or semantic parser could change relative scores, especially for rare classes.
Because few-shot LLMs perform poorly on rare classes (e.g., police-inform, general-greet), a testable extension is to add a few examples of these classes to the prompt or to use a retrieval-based example selection that covers rare intents.
The large gap between supervised and few-shot suggests that if a small LLM were fine-tuned on the same training data, it might outperform BERT, but the paper plans that as future work; no claim is made here.

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.

Desk Editor's Note

read the letter

You should know this paper is a straightforward, honest empirical comparison: three ~7B open-source LLMs do few-shot multi-label intent classification on MultiWOZ 2.1, and a small supervised BERT classifier beats them all. The headline result, that Mistral-7B-v0.1 gets a weighted F1 of 0.50 versus Llama-2's 0.40 and Yi's 0.29, is plausible and consistent with what I'd expect. The paper does a few things well: it reports per-class metrics, inference time, and VRAM usage; it includes a supervised baseline rather than hand-waving at the literature; and it openly discusses the risk of training-data contamination. That is a solid, useful contribution for practitioners choosing among these models on limited hardware.

The soft spots are real but not load-bearing. The biggest problem is that the prompt template is explicitly withheld (Section 4.2). The authors say it is 'relatively long' and do not include it. Since the ranking could easily depend on how each model responds to that specific prompt's wording and formatting, this is a reproducibility blocker. The strict parser (Section 4.3), which turns any non-JSON or unrecognized output into an empty list, could also penalize models that express intents correctly but in a slightly different surface form. That second concern is real but less severe: it affects absolute scores, and the relative ordering might survive a more lenient parser, but we don't know because raw outputs aren't included.

The numerical inconsistencies are minor but worth naming. Table 1 says 7,374 user turns, while the text and the support sums in Table 3 say 7,282. There's also a Jaccard typo in Section 5.2: the table says 0.379 but the text says 0.397. Those are easily fixed. The lack of repeated runs or confidence intervals is a more substantive limitation for a ranking claim, since a few hundred test samples per class and a single run mean the Mistral advantage might be noise. I'd like to see at least a prompt-variant check.

None of this makes the central argument wrong. The supervised BERT result is clean and matches the broader literature; the few-shot numbers are not crazy. If the authors release the prompt, the parser, and the raw outputs, the paper becomes genuinely useful. As written, it deserves peer review, but the referee should insist on that transparency before publication. I'd send it to a workshop or a systems-oriented venue rather than a top-tier conference, but it is not a desk reject.

Who is this for? Practitioners evaluating open-weight models for dialogue NLU on consumer GPUs. It is not a methods paper. If you work in that space, it's worth a skim; otherwise, the takeaway is the known result that supervised small models still win when you have labeled data.

Referee Report

4 major / 5 minor

Summary. The paper reports a few-shot multi-label intent classification comparison of three open-source 7B-level LLMs (Llama-2-7B-hf, Mistral-7B-v0.1, Yi-6B) on MultiWOZ 2.1 user utterances, using a single 20-example prompt selected on validation data. It also trains a BertForSequenceClassification baseline with and without previous system utterance context. The main empirical claims are: Mistral achieves the best few-shot weighted-average F-Score (0.50) and outperforms the other LLMs in 11 of 14 intent classes; and the supervised BERT baseline (0.92 micro/weighted F1 with context) substantially outperforms the few-shot LLMs. The paper also reports inference time, VRAM, Hamming loss, and Jaccard similarity.

Significance. If the comparison is reproducible, the paper provides a useful practical data point for practitioners choosing small open-source models for multi-intent dialogue NLU on a standard benchmark, with resource cost information and a supervised baseline that contextualizes the few-shot results. The BERT baseline is a strong point because it grounds the LLM numbers against a standard supervised approach. The paper's value is mainly empirical; the absence of a released prompt and raw outputs currently undermines the reproducibility that such a comparison needs.

major comments (4)

[§4.2, §5.2] The complete few-shot prompt is withheld ('As the prompt template is relatively long, we do not provide this in the paper content'), even though it is the central experimental instrument and was iteratively tuned on the validation set. Few-shot LLM results are highly sensitive to prompt wording, example selection, and ordering; using one undisclosed prompt across all three models means the ranking in Table 3—Mistral 11/14, weighted F1 0.50 vs 0.40 and 0.29—may reflect prompt compatibility rather than model capability. Please publish the full prompt and examples (appendix or repository) and provide at least a robustness check with alternative prompt formulations, or report per-model validation performance under the chosen prompt.
[§4.3] The parser discards any output not exactly matching the predefined intent set and returns an empty list. This penalizes models that express the same intent with different surface forms or valid non-JSON formatting. Because raw model outputs are not included, the reported F-scores cannot be audited or re-parsed under a more tolerant normalization. Please release raw outputs for all three models, report the fraction of empty/unparsed predictions per model, and show whether an alias-based or fuzzy normalization changes the ranking.
[Table 1 vs Table 3, §3, §5.3] The paper inconsistently reports the test set size: Table 1 says 7,374 user turns; §3 and §5.3 say 7,282; Table 3 support totals sum to 7,282. The difference of 92 turns may be due to parsing failures or excluded examples, but this is not stated. Please clarify the actual test set, the exclusion criterion, and its effect on the metrics; also align the Jaccard value (0.397 in §5.2 vs 0.379 in Table 3).
[§4.2, §5.3] The BERT decision threshold is chosen on the validation set from 9 candidates, and the prompt examples are chosen by validation experiments, but no sensitivity analysis or confidence intervals are reported. For an empirical ranking claim, it would be helpful to report variance across repeated BERT training runs (which has stochastic elements) or at least to show that the LLM ranking is unchanged under small perturbations to the prompt/parsing choices.

minor comments (5)

[Abstract/§5.2] 'Humming Loss' should be 'Hamming Loss'; the text's 'Jaccard Similarity of 0.397' conflicts with Table 3's 0.379. Please correct the value.
[§3] There is a duplicated fragment: 'the training data for supervised training.' appears after the sentence about using complete training/validation/test data. Remove the fragment or integrate it into a proper sentence.
[Table 3] Several score triples are concatenated without spaces (e.g., '0.440.55'), making the table hard to read. Use proper alignment and spacing in the table format.
[§4.1, footnote 7] The RTX A6000 is not a consumer-grade GPU; this characterization is inaccurate. Also, footnote 7 links to Llama-2-70b-chat-hf, not the 7B model used in the experiments.
[Throughout] The abbreviation 'somoLLMs' is used without definition; define it at first use. Model-name capitalization is inconsistent ('LLama2-7B-hf' vs 'Llama-2-7b-hf' vs 'Llama-2-7B-hf').

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical benchmark with held-out test evaluation; validation-based prompt/threshold tuning is standard model selection, not a fitted prediction.

full rationale

The paper is an empirical comparison, not a derivation. The central claim that Mistral-7B-v0.1 outperforms Llama-2-7B and Yi-6B on few-shot multi-intent classification is evaluated against the MultiWOZ 2.1 test set using ground-truth intent labels. The prompt is selected using the validation split (Section 4.2: 'we use that prompt across different models to perform prediction on the test data'), and the BERT threshold is also chosen on the validation set (Section 5.3), before being applied to the test set. This is standard model selection, not fitting a parameter to the target result and then calling it a prediction. No equation in the paper defines the reported F-scores in terms of the fitted prompt or threshold; the test set is independent. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The main weaknesses—the undisclosed prompt template, the strict parser that discards off-format intents, and the slight inconsistencies in turn counts and Jaccard values—are reproducibility and evaluation-validity concerns, not circularity. They may affect how much faith one places in the ranking, but they do not make the reported result equivalent to its inputs by construction. Therefore the circularity score is 0.

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

The paper introduces no new mathematical objects. The main load-bearing assumptions are the validity of the dataset labels, the chosen parsing rule for model outputs, and the transferability of one validation-tuned prompt to all models. Free parameters are the BERT threshold and the few-shot prompt content, both tuned on validation.

free parameters (3)

BERT decision threshold = 0.5
Threshold selected by scanning 9 candidate values and micro F1 on the validation set (Section 5.3).
Few-shot prompt examples = 20 examples, content tuned on validation results
Examples and instruction wording were added iteratively based on poor-performing classes on a portion of validation data (Section 4.2).
Sampling temperature = 0
Set to 0 for deterministic outputs (Section 4.3). Chosen by hand, not fitted, but it affects output variety.

assumptions (3)

domain assumption MultiWOZ 2.1 annotations are the correct ground truth for intents
All evaluation metrics are computed against these labels (Sections 3 and 5.1).
domain assumption LLM outputs with no recognized intent are treated as an empty prediction
The parser returns an empty list if the model hallucinates an intent or produces no intents (Section 4.3). This scoring rule may penalize differently formatted outputs.
domain assumption The prompt tuned on validation is equally effective across all three models
The same prompt was used across models (Section 4.2, 5.2); model-specific prompt sensitivity is not tested.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Intent Recognition in Dialogue Understanding: A Comparison Between Smaller Open-Source LLMs." pith.science (2026). https://pith.science/paper/TEVQFOXG

@misc{pith2026250910010,
  author       = {Pith},
  title        = {Pith review of: Multi-Intent Recognition in Dialogue Understanding: A Comparison Between Smaller Open-Source LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TEVQFOXG}},
  note         = {Machine review of arXiv:2509.10010}
}

read the original abstract

In this paper, we provide an extensive analysis of multi-label intent classification using Large Language Models (LLMs) that are open-source, publicly available, and can be run in consumer hardware. We use the MultiWOZ 2.1 dataset, a benchmark in the dialogue system domain, to investigate the efficacy of three popular open-source pre-trained LLMs, namely LLama2-7B-hf, Mistral-7B-v0.1, and Yi-6B. We perform the classification task in a few-shot setup, giving 20 examples in the prompt with some instructions. Our approach focuses on the differences in performance of these models across several performance metrics by methodically assessing these models on multi-label intent classification tasks. Additionally, we compare the performance of the instruction-based fine-tuning approach with supervised learning using the smaller transformer model BertForSequenceClassification as a baseline. To evaluate the performance of the models, we use evaluation metrics like accuracy, precision, and recall as well as micro, macro, and weighted F1 score. We also report the inference time, VRAM requirements, etc. The Mistral-7B-v0.1 outperforms two other generative models on 11 intent classes out of 14 in terms of F-Score, with a weighted average of 0.50. It also has relatively lower Humming Loss and higher Jaccard Similarity, making it the winning model in the few-shot setting. We find BERT based supervised classifier having superior performance compared to the best performing few-shot generative LLM. The study provides a framework for small open-source LLMs in detecting complex multi-intent dialogues, enhancing the Natural Language Understanding aspect of task-oriented chatbots.

Figures

Figures reproduced from arXiv: 2509.10010 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 12 linked inside Pith

[1]

and Agarwal, S., 2020

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A. and Agarwal, S., 2020. Language models are few-shot learn- ers. Advances in neural information processing systems, 33, pp.1877-1901

2020
[2]

and Schuh, P., 2022

Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S. and Schuh, P., 2022. Palm: Scaling language mod- eling with pathways. arXiv preprint arXiv:2204.02311

arXiv 2022
[3]

and Le, Q.V .,

Wei, J., Bosma, M., Zhao, V .Y ., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M. and Le, Q.V .,
[4]

and Webson, A., 2022

Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y ., Fedus, W., Li, Y ., Wang, X., Dehghani, M., Brahma, S. and Webson, A., 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416

arXiv 2022
[5]

and Schulman, J., 2022

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A. and Schulman, J., 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, pp.27730- 27744

2022
[6]

Anthropic Blog 2023, March 14

Anthropic Introducing Claude. Anthropic Blog 2023, March 14. https://www.anthropic.com/index/introducing-claude (accessed on 19 November 2023)

2023
[7]

Anthropic Blog 2023, July 11

Anthropic Claude 2. Anthropic Blog 2023, July 11. https://www.anthropic.com/index/claude- 2 (accessed on 19 November 2023)

2023
[8]

and Bahdanau, D., 2022

Sahu, G., Rodriguez, P., Laradji, I.H., Atighehchian, P., Vazquez, D. and Bahdanau, D., 2022. Data augmentation for intent classification with off-the-shelf large language models. arXiv preprint arXiv:2204.01959

arXiv 2022

Show all 27 references

[9]

and Bikel, D., 2023

Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y ., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S. and Bikel, D., 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

2023 arXiv
[10]

and Lavaud, L.R., 2023

Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.D.L., Bres- sand, F., Lengyel, G., Lample, G., Saulnier, L. and Lavaud, L.R., 2023. Mistral 7B. arXiv preprint arXiv:2310.06825. 12 Adnan Ahmad and Philine Kowol and Stefan Hillmann and Sebastian M ¨oller

2023 arXiv
[11]

https://github.com/01-ai/Yi (accessed on 19 November 2023)

Yi model github repository, November 2023. https://github.com/01-ai/Yi (accessed on 19 November 2023)

2023
[12]

and Li, X., 2023

Hao, Z., Zhang, J., Xu, B., Wang, Y ., de Melo, G. and Li, X., 2023. IntentDial: An Intent Graph based Multi-Turn Dialogue System with Reasoning Path Visualization. arXiv preprint arXiv:2310.11818

2023 arXiv
[13]

and Xiong, C.,

Li, S., Yavuz, S., Hashimoto, K., Li, J., Niu, T., Rajani, N., Yan, X., Zhou, Y . and Xiong, C.,
[14]

and Awadallah, A.H., 2020, October

Yu, T., Zhang, R., Polozov, A., Meek, C. and Awadallah, A.H., 2020, October. Score: Pre- training for context representation in conversational semantic parsing. In International Con- ference on Learning Representations

2020
[15]

and Hakkani-Tur, D., 2019

Eric, M., Goel, R., Paul, S., Kumar, A., Sethi, A., Ku, P., Goyal, A.K., Agarwal, S., Gao, S. and Hakkani-Tur, D., 2019. MultiWOZ 2.1: A consolidated multi-domain dialogue dataset with state corrections and state tracking baselines. arXiv preprint arXiv:1907.01669

2019 arXiv
[16]

& Wolf, T., 2023

Beeching, E., Fourrier, C., Habib, N., Han, S., Lambert, N., Rajani, N., Sanseviero, O., Tunstall, L. & Wolf, T., 2023. Open LLM Leaderboard. Hugging Face. Avail- able at: https://huggingface.co/spaces/HuggingFaceH4/open llm leaderboard (accessed on 19 November 2023)

2023
[17]

Budzianowski, P., et al. (2018). ”MultiWOZ - A Large-Scale Multi-Domain Wizard-of-Oz Dataset for Task-Oriented Dialogue Modelling.” In Proceedings of the EMNLP

2018
[18]

Rastogi, A., et al. (2020). ”Towards Scalable Multi-domain Conversational Agents: The Schema-Guided Dialogue Dataset.” In AAAI

2020
[19]

and Huang, M., 2021

Han, T., Liu, X., Takanabu, R., Lian, Y ., Huang, C., Wan, D., Peng, W. and Huang, M., 2021. Multiwoz 2.3: A multi-domain task-oriented dialogue dataset enhanced with annotation cor- rections and co-reference annotation. In Natural Language Processing and Chinese Comput- ing: ...

2021
[20]

and Chen, J., 2020

Zang, X., Rastogi, A., Sunkara, S., Gupta, R., Zhang, J. and Chen, J., 2020. MultiWOZ 2.2: A dialogue dataset with additional annotation corrections and state tracking baselines. arXiv preprint arXiv:2007.12720

2020 arXiv
[21]

and Yilmaz, E., 2021

Ye, F., Manotumruksa, J. and Yilmaz, E., 2021. Multiwoz 2.4: A multi-domain task-oriented dialogue dataset with essential annotation corrections to improve state tracking evaluation. arXiv preprint arXiv:2104.00773

2021 arXiv
[22]

and Cree, M., 2016, November

Gouk, H., Pfahringer, B. and Cree, M., 2016, November. Learning distance metrics for multi- label classification. In Asian Conference on machine learning (pp. 318-333). PMLR

2016
[23]

and Zhu, X., 2021

Dai, Y ., Li, H., Li, Y ., Sun, J., Huang, F., Si, L. and Zhu, X., 2021. Preview, attend and review: Schema-aware curriculum learning for multi-domain dialog state tracking. arXiv preprint arXiv:2106.00291

2021 arXiv
[24]

(2019).BERT: Pre-training of Deep Bidi- rectional Transformers for Language Understanding

Devlin, J., Chang, M.-W., Lee, K., Toutanova, K. (2019).BERT: Pre-training of Deep Bidi- rectional Transformers for Language Understanding. InProc. NAACL 2019, ACL, pp. 4171–

2019
[2020]

arXiv preprint arXiv:2010.12850

Coco: Controllable counterfactuals for evaluating dialogue state trackers. arXiv preprint arXiv:2010.12850

2010 arXiv
[2021]

arXiv preprint arXiv:2109.01652

Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652

arXiv
[4186]

doi: 10.18653/v1/N19-1423

doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

and Agarwal, S., 2020

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A. and Agarwal, S., 2020. Language models are few-shot learn- ers. Advances in neural information processing systems, 33, pp.1877-1901

2020

[2] [2]

and Schuh, P., 2022

Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S. and Schuh, P., 2022. Palm: Scaling language mod- eling with pathways. arXiv preprint arXiv:2204.02311

arXiv 2022

[3] [3]

and Le, Q.V .,

Wei, J., Bosma, M., Zhao, V .Y ., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M. and Le, Q.V .,

[4] [4]

and Webson, A., 2022

Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y ., Fedus, W., Li, Y ., Wang, X., Dehghani, M., Brahma, S. and Webson, A., 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416

arXiv 2022

[5] [5]

and Schulman, J., 2022

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A. and Schulman, J., 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, pp.27730- 27744

2022

[6] [6]

Anthropic Blog 2023, March 14

Anthropic Introducing Claude. Anthropic Blog 2023, March 14. https://www.anthropic.com/index/introducing-claude (accessed on 19 November 2023)

2023

[7] [7]

Anthropic Blog 2023, July 11

Anthropic Claude 2. Anthropic Blog 2023, July 11. https://www.anthropic.com/index/claude- 2 (accessed on 19 November 2023)

2023

[8] [8]

and Bahdanau, D., 2022

Sahu, G., Rodriguez, P., Laradji, I.H., Atighehchian, P., Vazquez, D. and Bahdanau, D., 2022. Data augmentation for intent classification with off-the-shelf large language models. arXiv preprint arXiv:2204.01959

arXiv 2022