Datasets:

jamescalam

/

ai-arxiv-llama2-ja

like 0

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\n33B 0.23 0.26 0.34 0.50 0.36 0.35 0.24 0.33 0.34 0.49 0.31 0.12 0.23 0.30 0.41 0.60 0.28 0.27\n65B 0.25 0.26 0.34 0.46 0.36 0.40 0.25 0.32 0.32 0.48 0.31 0.11 0.25 0.30 0.43 0.60 0.39 0.34\nL/l.sc/a.sc/m.sc/a.sc /two.taboldstyle7B 0.28 0.25 0.29 0.50 0.36 0.37 0.21 0.34 0.32 0.50 0.28 0.19 0.26 0.32 0.44 0.51 0.30 0.25\n13B 0.24 0.25 0.35 0.50 0.41 0.36 0.24 0.39 0.35 0.48 0.31 0.18 0.27 0.34 0.46 0.66 0.35 0.28" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\n34B 0.27 0.24 0.33 0.56 0.41 0.36 0.26 0.32 0.36 0.53 0.33 0.07 0.26 0.30 0.45 0.56 0.26 0.35\n70B 0.31 0.29 0.35 0.51 0.41 0.45 0.27 0.34 0.40 0.52 0.36 0.12 0.28 0.31 0.45 0.65 0.33 0.20\nFine-tuned\nChatGPT 0.65 0.62 0.64 0.84 0.77 0.75 0.53 0.71 0.73 0.75 0.73 0.54 0.55 0.69 0.71 0.82 0.57 0.57\nMPT-instruct 7B 0.22 0.19 0.28 0.44 0.27 0.26 0.19 0.28 0.30 0.46 0.24 0.05 0.20 0.39 0.33 0.48 0.20 0.19\nFalcon-instruct 7B 0.36 0.31 0.48 0.62 0.48 0.45 0.31 0.47 0.40 0.57 0.43 0.19 0.30 0.56 0.47 0.63 0.49 0.48" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nL/l.sc/a.sc/m.sc/a.sc /two.taboldstyle-C/h.sc/a.sc/t.sc7B 0.44 0.42 0.45 0.71 0.54 0.54 0.33 0.54 0.53 0.55 0.62 0.29 0.36 0.58 0.53 0.61 0.36 0.37\n13B 0.37 0.37 0.41 0.52 0.44 0.45 0.29 0.46 0.49 0.50 0.48 0.29 0.31 0.58 0.41 0.58 0.33 0.40\n34B 0.40 0.37 0.43 0.59 0.54 0.49 0.32 0.48 0.50 0.58 0.53 0.25 0.34 0.60 0.50 0.63 0.44 0.40\n70B 0.47 0.43 0.49 0.67 0.60 0.55 0.38 0.54 0.56 0.61 0.58 0.28 0.39 0.67 0.56 0.70 0.43 0.47\nTable50: Distribution of mean sentiment scores across groups under the profession domain from the BOLD\nprompts.\nA.5.2 Negative User Experience Categories" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nTable50: Distribution of mean sentiment scores across groups under the profession domain from the BOLD\nprompts.\nA.5.2 Negative User Experience Categories\nThere are different kinds of responses that could cause a negative user experience when interacting with our\nmodels. We have instructed the annotators to avoid writing responses that violate our safety guidelines, for\nexample, we ask that prompts they write do not:\n1. Promote or enable criminal activities.\n2. Promote or enable dangerous behaviors to the user or other people.\n3. Contain, promote or enable offensive and abusive behavior towards the user or other people.\n4. Contain, promote or enable sexually explicit content.\nA.5.3 Quality Assurance Process\nWehaveimplementedaqualityassuranceprocesstoensureweonlyusehighqualityannotationsfortraining\nthemodel. Forthisprocess,ateamofhighlyskilledcontentmanagersmanuallyreviewedtheannotations\nand approved the ones that would be used.\nDuringthequalityassurance step, reviewerswere asked toonly approve those annotations thatmatched our\nguidelines: (a)theyareconsistentwiththedialoguehistory,(b)followinstructionsintheprompt(c)are\nfreeofgrammatical,spellingandotherwritingerrors,and(d)donotfallintoanyofthecategoriesdescribed" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nfreeofgrammatical,spellingandotherwritingerrors,and(d)donotfallintoanyofthecategoriesdescribed\ninSectionA.5.2. Ifanannotationneededsmallchangestobeapproved,duetogrammarorspellingmistakes,\nortoimprovethestructure,cohesivenessandstyleofthetext,reviewerscouldeditittofixtheissuesand\napprove it. If the answer could not be approved without major changes, the reviewers were asked to reject it\nand write the feedback necessary to improve it.\nA.5.4 Annotator Selection\nToselecttheannotatorswhocouldworkonourdifferentdatacollectiontasks,weconductedamulti-step\nassessment process where we tested their understanding of our guidelines, the alignment with our quality\nassessment criteria, the alignment with our sensitive topics guidelines and their reading and writing skills.\nThe process included 4 tests:\n•Thefirsttestconsistsof3sectionsoftestingtoevaluategrammar,readingcomprehensionandwriting\nstyle. Each section is timed andthe test should take atotalof 50 minutes to complete. A candidate\nmustscore90%onpartItocontinueontopartsIIandIII,andanaveragescoreof4onpartIIandIII\nto pass the test." }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nmustscore90%onpartItocontinueontopartsIIandIII,andanaveragescoreof4onpartIIandIII\nto pass the test.\n•Thesecondtestconsistedof42questionssplitintosensitivetopicsalignment,answerrankingand\ntwo examples of answer writing, which were manually reviewed by us. To pass the test, annotators\nneeded to agree with our criteria on 80% of the answers, and pass the written examples with a score\nof 4 out of 5.\n74\n•Thethirdtestconsistedinmeasuringthealignmentwithourqualityassessmentcriteria. Thetest\nconsisted of 31 different questions asking the annotators to grade different prompt-answer pairs,\naswellasrankingdifferentanswerstothesameprompt. Tomeasurealignment,wefirstcollected\nresponses from different team members, and the annotators who agreed with our preferences in\nmore than 26 of the questions passed the test.\n•Finally,thelasttestconsistedofapromptresponseassessmentwhereannotatorschooseaminimumof\n6outof18promptstowriteresponsesfor. Wemanuallyassesseachresponsetoevaluateproduction\nreadiness. Annotators that have scored an average of >4 have passed the training." }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nreadiness. Annotators that have scored an average of >4 have passed the training.\nA.6 Dataset Contamination\nWiththeincreasingscaleofpubliclyavailabletrainingdata,ithasbecomeinevitablethatsomeportionof\nevaluation data is seen during training, and may provide an undue boost in evaluation performance.\nEarlierwork(Brownetal.(2020),Weietal.(2022a),Duetal.(2022)inmeasuringsuchdatasetcontamination\nconsidered an example from an evaluation set to be “contaminated” if there existed a collision between\na high-order n-gram (generally, n= 13) from the sample and the training data. This was a deliberately\nconservativeapproachinordertoproducea“clean”subsetofthedatawithhighprecision,andisusedin\nopen-sourced evaluation libraries (e.g. Gao et al. (2021)).\nThis approach, however, was unable to detect precisely what proportion of a given sample is contaminated,\nand didn’t take into account how evaluation datasets are constructed. Furthermore, as noted in Chowdhery\netal.(2022),somedatasets(suchasBoolQ)containcontextsextractedverbatimfromtheweb,butnotthe" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\netal.(2022),somedatasets(suchasBoolQ)containcontextsextractedverbatimfromtheweb,butnotthe\nquestionandanswercontinuation. Assuch,highlycontaminatedsamplesfromthesedatasetsareunlikely\ntogainanunfairadvantage. ThemethodologyinChowdheryetal.(2022)furtherimprovesontheearlier\nn-gram collision detection by considering a sample to be contaminated if 70% of all 8-grams can be found at\nleast once in the training data.\nThe previous methodologies noted above all consider contamination in text space, and don’t appear to\nconsidertheformattingofpromptsusedforactualevaluation. Incontrast,weinsteadmatchontokenized\ninput,beingcarefultopassfullyverbalizedevaluationsamplestothetokenizer. Wealsodivergefromthe\nprevious methodologies by considering contamination from a bottom-up perspective. We consider a token\ntobecontaminatedifitappearsinanytoken n-gramlongerthan10tokensinboththeevaluationsample\nand the training set, and define the contamination percentage of a sample to be the percentage of tokens" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nand the training set, and define the contamination percentage of a sample to be the percentage of tokens\ncontaminated. Thisallowsustoviewthebenchmarkperformanceofourmodelsonarangeofcontamination\nscales, while retaining the ability to test a high-precision clean subset (samples with <20%contamination)\nand a high-precision contaminated subset (samples with >80%contamination). In order to account for the\nvagaries of the precise format of verbalized samples, we allow a small \"skipgram budget\" of four tokens, so\nthat matchedspans between anevaluation sampleand the trainingdata can differin at mostfour positions\n(we do not allow trailing mismatches, or mismatches in the first 10 tokens).\nWe identify such 10(+)-skipgrams with suffix arrays implemented using a variation of the library from Lee\netal.(2022),modifiedtoworkonaPySparkcluster(effectivelywithoutrandomaccesstodisk). Giventhe\nembarrassinglyparallelnatureofthetask,weareabletofindallsuch10-grams(andtheirfulllengths)in\nour entire dataset in around seven hours (including time to tokenize), utilizing an estimated 1,500 cores." }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nour entire dataset in around seven hours (including time to tokenize), utilizing an estimated 1,500 cores.\nAs there are many confounding factors at play when determining whether dataset contamination has\ncontributedtoevaluationperformance(mostlystemmingfromthefactthat\"clean\"and\"dirty\"subsetsdo\nnot necessarily well-estimate the population distribution), we make the following assumption: In the event\nof dataset contamination contributing to evaluation performance, we expect both the \"cleanest\" examples to\nhaveanoverall worseaveragescorethantheircomplement,andthe\"dirtiest\"samplestohaveanoverall better\naverage score than their complement. It is insufficient evidence for contamination if only one of these were\ntrue. To this end, we define four (non-disjoint) subset types as follows:\n•“Clean”samples, with less than 20% token contamination,\n•“Not clean” samples, with greater than (or equal to) 20% token contamination,\n•“Not dirty” samples, with less than 80% token contamination,\n•“Dirty”samples, with greater than (or equal to) 80% token contamination.\nThereisanadditionalconfoundingfactorthatweattempttoaddressdirectly. Withthegivendefinitionof" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nThereisanadditionalconfoundingfactorthatweattempttoaddressdirectly. Withthegivendefinitionof\ncontamination(aswellasotherdefinitionsmentionedintheliterature),thereisapossibilitythatasample\n75\nDataset Model Subset Type Avg. Contam. % n \u0016X \u0016nZn\nHellaSwag ( L= 40)70BClean 0 7391 80.0 82.5 -5.73\nNot Clean 67.5 2651 89.5 82.4 9.56\nNot Dirty 11.5 9194 81.6 82.5 -2.27\nDirty 86.1 848 92.2 82.5 7.42\n7BClean 0 7391 70.5 73.3 -5.46\nNot Clean 67.5 2651 81.3 73.4 9.17\nNot Dirty 11.5 9194 72.4 73.4 -2.06\nDirty 86.1 848 83.7 73.3 6.84\nMMLU-Humanities ( L= 50)70BClean 0.05 3996 62.2 65.3 -4.08\nNot Clean 85.12 709 82.7 65.3 9.71\nNot Dirty 2.73 4185 62.7 65.3 -3.50" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nNot Clean 85.12 709 82.7 65.3 9.71\nNot Dirty 2.73 4185 62.7 65.3 -3.50\nDirty 94.5 520 85.8 65.3 9.80\n7BClean 0.05 3996 40.8 42.9 -2.75\nNot Clean 85.2 709 54.9 42.8 6.50\nNot Dirty 2.73 4185 41.1 42.9 -2.25\nDirty 94.5 520 56.9 42.8 6.49\nMMLU-Overall ( L= 50) 70BClean 0.02 11862 68.0 68.9 -2.00\nNot Clean 84.7 2180 73.5 68.9 4.64\nNot Dirty 3.18 12506 67.7 68.9 -2.75\nDirty 94.4 1536 78.2 68.9 7.87\nTable 51: Contamination analysis results for affected datasets. No other evaluation datasets had sufficient\nevidence to be considered affected by contamination. Avg. Contam. % denotes the average per-sample\ncontamination percentage for the given subset type. Models sizes refer to pretrained-only models\nmay appear contaminated, by virtue of many tokens appearing in matched sequences found in the training" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\ncontamination percentage for the given subset type. Models sizes refer to pretrained-only models\nmay appear contaminated, by virtue of many tokens appearing in matched sequences found in the training\ndata. However, the matched sequences might be highly fragmented across the training data, in which case it\nis very unlikely the model saw the correctly-assembled contaminated sequences during training. To reduce\nthe chance of this phenomenon, we repeat our analysis with minimum match length L2f10;20;30;40;50g.\nSinceinthelimitof L!1everysamplefallsintoboththe\"clean\"and\"notdirty\"(thereisnocontamination),\nwe report the largest Lfor each dataset that appeared to benefit from contamination to strike a balance\nbetween fragmentation and overall contamination.\nForeachdatasetandeachoftheabovesamplesubsettypes,wecomputeboththemean \u0016Xoftheperformance\nmetricXand the statistic Zn=(\u0016X\u0000\u0016n)\n\u001bn, wherenis the size of the sample subset type, and \u0016nand\u001b2\nnare the\nmean and variance of the sampling distribution of the performance metric for samples of size n, respectively.\nBytheCentralLimitTheorem, Zntendstowardsastandardnormaldistributionandsoweconsiderthere" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nBytheCentralLimitTheorem, Zntendstowardsastandardnormaldistributionandsoweconsiderthere\nissufficientevidencetosuggestcontaminationhasaffectedevaluationperformanceonadatasetifallfour\nsample subsets have jZnj>2.\nResultsforthisanalysiscanbeseeninTable51. WeobservethatonlyHellaSwagandMMLU-Humanities\nappear to have been boosted due to contamination in the training data, with the 70B model appearing to\nhave gained a greater benefit than the 7B model, as one might expect. Furthermore, the impact of this effect\nonMMLU-HumanitiesappearstocauseabenefitforMMLU-Overallforthe70Bmodel,albeitwithonlya\nsmall delta (-0.9) between the \"clean\" subset performance and the sampling mean. No other dataset (for any\nchoice ofL) appears to have benefitted from dataset contamination, and we omit results from these datasets\nfor conciseness.\n76\nA.7 Model Card\nTable 52 presents a model card (Mitchell et al., 2018; Anil et al., 2023) that summarizes details of the models.\nModel Details\nModel Developers Meta AI" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nTable 52 presents a model card (Mitchell et al., 2018; Anil et al., 2023) that summarizes details of the models.\nModel Details\nModel Developers Meta AI\nVariations L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle comes in a range of parameter sizes—7B, 13B, and 70B—as well as\npretrained and fine-tuned variations.\nInput Models input text only.\nOutput Models generate text only.\nModel Architecture L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle isanauto-regressivelanguagemodelthatusesanoptimizedtransformer\narchitecture. Thetunedversionsusesupervisedfine-tuning(SFT)andreinforcementlearning withhuman feedback(RLHF)to aligntohuman preferencesfor\nhelpfulness and safety.\nModel Dates L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle was trained between January 2023 and July 2023.\nStatus This is a static model trained on an offline dataset. Future versions of the tuned\nmodels will be released as we improve model safety with community feedback.\nLicense A custom commercial license is available at: ai.meta.com/resources/" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nmodels will be released as we improve model safety with community feedback.\nLicense A custom commercial license is available at: ai.meta.com/resources/\nmodels-and-libraries/llama-downloads/\nWhere to send commentsInstructions on how to provide feedback or comments on the model can be\nfound in the model README, or by opening an issue in the GitHub repository\n(https://github.com/facebookresearch/llama/ ).\nIntended Use\nIntended Use Cases L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle is intended for commercial and research use in English. Tuned models\nare intended for assistant-like chat, whereas pretrained models can be adapted\nfor a variety of natural language generation tasks.\nOut-of-Scope Uses Use in any manner that violates applicable laws or regulations (including trade\ncompliancelaws). UseinlanguagesotherthanEnglish. Useinanyotherway\nthat is prohibited by the Acceptable Use Policy and Licensing Agreement for\nL/l.sc/a.sc/m.sc/a.sc /two.taboldstyle.\nHardware and Software (Section 2.2)\nTraining Factors We usedcustomtraininglibraries, Meta’sResearchSuperCluster, andproductionclustersforpretraining. Fine-tuning,annotation,andevaluationwerealso" }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nTraining Factors We usedcustomtraininglibraries, Meta’sResearchSuperCluster, andproductionclustersforpretraining. Fine-tuning,annotation,andevaluationwerealso\nperformed on third-party cloud compute.\nCarbon Footprint Pretraining utilized a cumulative 3.3M GPU hours of computation on hardware\nof type A100-80GB (TDP of 350-400W). Estimated total emissions were 539\ntCO 2eq, 100% of which were offset by Meta’s sustainability program.\nTraining Data (Sections 2.1 and 3)\nOverview L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle was pretrained on 2 trillion tokens of data from publicly available\nsources. The fine-tuning data includes publicly available instruction datasets, as\nwellasoveronemillionnewhuman-annotatedexamples. Neitherthepretraining\nnor the fine-tuning datasets include Meta user data.\nData Freshness The pretraining data has a cutoff of September 2022, but some tuning data is\nmore recent, up to July 2023.\nEvaluation Results\nSee evaluations for pretraining (Section 2); fine-tuning (Section 3); and safety (Section 4)." }

{ "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models", "en": "Llama 2: Open Foundation and Fine-Tuned Chat Models\nEvaluation Results\nSee evaluations for pretraining (Section 2); fine-tuning (Section 3); and safety (Section 4).\nEthical Considerations and Limitations (Section 5.2)\nL/l.sc/a.sc/m.sc/a.sc /two.taboldstyle is a new technology that carries risks with use. Testing conducted to date has been in\nEnglish, and has notcovered, nor could it coverall scenarios. For these reasons, aswith all LLMs,\nL/l.sc/a.sc/m.sc/a.sc /two.taboldstyle’s potential outputs cannot be predicted in advance, and the model may in some instances\nproduceinaccurateorobjectionableresponsestouserprompts. Therefore,beforedeployingany\napplications of L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle, developers should perform safety testing and tuning tailored to their\nspecific applications of the model. Please see the Responsible Use Guide available available at\nhttps://ai.meta.com/llama/responsible-user-guide\nTable 52: Model card for L/l.sc/a.sc/m.sc/a.sc /two.taboldstyle .\n77" }