The quality and structure of training data matters more than its volume.
这一观点强调了数据质量在模型训练中的重要性，为数据工程和模型训练提供了新的方向。
32 Matching Annotations
- Last 7 days
-
www.llmwatch.com www.llmwatch.com
-
- Apr 2026
-
x.com x.com
-
frontier AI models are not too big because the technology is complex and too big because the training data is garbage
这一观点挑战了当前AI模型规模扩大的主流解释，将问题从技术复杂性转向数据质量问题，提出了一个反直觉的视角：模型规模实际上是应对低质量数据的必要之举，而非技术发展的必然结果。
-
-
-
A small model trained on fewer than 2,000 examples from real lawyers, bankers, and consultants recently beat all but the best frontier models on corporate legal work, at a fraction of the price.
这一发现挑战了'规模和计算能力胜过一切'的AI发展范式。高质量专业化数据训练的小型模型在特定领域表现优于通用大模型，暗示AI发展可能从'越大越好'转向'更专业、更高效'的新阶段。
-
-
-
As slop takes over the Internet, labs may struggle to obtain high-quality corpuses for training models.
这一观察揭示了AI训练数据质量的危机。随着互联网内容质量的下降，AI系统可能面临'垃圾进，垃圾出'的风险。作者提出的'低背景钢'比喻巧妙地指出了使用2023年前纯净数据的解决方案，同时也暗示了数字时代知识污染的严重性，这可能会对AI系统的可靠性和偏见产生深远影响。
-
-
huggingface.co huggingface.co
-
A three-stage progressive training strategy -- large-scale pre-training, hard sample fine-tuning, and GRPO alignment -- sequentially exploits these data at different quality tiers.
大多数人认为训练策略应该统一应用于所有数据，但作者提出了分阶段渐进式训练策略，在不同质量层级的数据上采用不同方法，这种针对数据质量差异的训练方法挑战了传统'一刀切'的训练范式，代表了数据为中心的AI新思路。
-
- Mar 2026
-
www.imsglobal.org www.imsglobal.org
-
Comprehensive Learner Record Standard Transcript Guide
CLR Playbook
-
- Dec 2025
-
link.springer.com link.springer.com
-
Micro-Credentials and Digital Badges: An Exploration of Definitions and Implications in Higher Education and Workforce
-
- Apr 2022
-
twitter.com twitter.com
-
The Lancet. (2021, April 16). Quantity > quality? The magnitude of #COVID19 research of questionable methodological quality reveals an urgent need to optimise clinical trial research—But how? A new @LancetGH Series discusses challenges and solutions. Read https://t.co/z4SluR3yuh 1/5 https://t.co/94RRVT0qhF [Tweet]. @TheLancet. https://twitter.com/TheLancet/status/1383027527233515520
-
- Feb 2022
-
Local file Local file
-
The unified approachhas the advantage, that the enterprise has more control overthe data and quality, and the data querying is significantlyfaster.
-
- Jan 2022
-
virologydownunder.com virologydownunder.com
-
Mackay, A. I. M. & PhD. (2022, January 29). Thank goodness we did all the work. Virology Down Under. https://virologydownunder.com/thank-goodness-we-did-all-the-work/
-
-
hackmd.io hackmd.io
-
Argument quality and fallacies. (n.d.). HackMD. Retrieved January 17, 2022, from https://hackmd.io/@scibehC19vax/argumentquality
Tags
- argument quality
- lang:en
- Simpson's paradox
- vaccine hesitancy
- ignorance
- arguments
- evidence
- vaccination debate
- norms
- source reliability
- factual error
- standards
- bias
- self-contradiction
- is:article
- fallacies
- vaccine data
- claim
- ad hominem argument
- causation
- statistical fallacies
- inconsistency
- slippery slope
Annotators
URL
-
- Jun 2021
-
psyarxiv.com psyarxiv.com
-
Moss, A. J., Rosenzweig, C., Jaffe, S. N., Gautam, R., Robinson, J., & Litman, L. (2021). Bots or inattentive humans? Identifying sources of low-quality data in online platforms [Preprint]. PsyArXiv. https://doi.org/10.31234/osf.io/wr8ds
-
- Feb 2021
-
psyarxiv.com psyarxiv.com
-
Lo, C., Mani, N., Kartushina, N., Mayor, J., & Hermes, J. (2021, February 11). e-Babylab: An Open-source Browser-based Tool for Unmoderated Online Developmental Studies. https://doi.org/10.31234/osf.io/u73sy
-
- Nov 2020
-
psyarxiv.com psyarxiv.com
-
Bauer, B., Larsen, K. L., Caulfield, N., Elder, D., Jordan, S., & Capron, D. (2020). Review of Best Practice Recommendations for Ensuring High Quality Data with Amazon’s Mechanical Turk. PsyArXiv. https://doi.org/10.31234/osf.io/m78sf
-
- Oct 2020
-
ecfr.federalregister.gov ecfr.federalregister.gov
-
(d) All calculations shown in this appendix shall be implemented on a site-level basis. Site level concentration data shall be processed as follows: (1) The default dataset for PM2.5 mass concentrations for a site shall consist of the measured concentrations recorded from the designated primary monitor(s). All daily values produced by the primary monitor are considered part of the site record; this includes all creditable samples and all extra samples. (2) Data for the primary monitors shall be augmented as much as possible with data from collocated monitors. If a valid daily value is not produced by the primary monitor for a particular day (scheduled or otherwise), but a value is available from a collocated monitor, then that collocated value shall be considered part of the combined site data record. If more than one collocated daily value is available, the average of those valid collocated values shall be used as the daily value. The data record resulting from this procedure is referred to as the “combined site data record.”
- Calculate mean of all collocated NON-primary monitors' values per day
- Coalesce primary monitor value with this calculated mean
-
-
aqs.epa.gov aqs.epa.gov
-
1.1. Monitors For the purposes of AQS, a monitor does not refer to a specific piece of equipment. Instead, it reflects that a given pollutant (or other parameter) is being measured at a given site. Identified by: The site (state + county + site number) where the monitor is located AND The pollutant code AND POC – Parameter Occurrence Code. Used to uniquely identify a monitor if there is more than one device measuring the same pollutant at the same site. For example monitor IDs are usually written in the following way: SS-CCC-NNNN-PPPPP-Q where SS is the State FIPS code, CCC is the County FIPS code, and NNNN is the Site Number within the county (leading zeroes are always included for these fields), PPPPP is the AQS 5-digit parameter code, and Q is the POC. For example: 01-089-0014-44201-2 is Alabama, Madison County, Site Number 14, ozone monitor, POC 2.
How monitors (specific measures of specific criteria) are identified in AQS data.
-
- Sep 2020
-
conbio.onlinelibrary.wiley.com conbio.onlinelibrary.wiley.com
-
Had it not been for the attentiveness of one person who went beyond the task of classifying galaxies into predetermined categories and was able to communicate this to the researchers via the online forum, what turned out to be important new phenomena might have gone undiscovered.
Sometimes our attempts to improve data quality in citizen science projects can actually work against us. Pre-determined categories and strict regulations could prevent the reporting of important outliers.
-
- Aug 2020
-
-
Harford, Tim. ‘A Bluffer’s Guide to Surviving Covid-19 | Free to Read’, 28 August 2020. https://www.ft.com/content/176b9bbe-56cf-4428-a0cd-070db2d8e6ff.
-
-
academic.oup.com academic.oup.com
-
van Smeden, M., Lash, T. L., & Groenwold, R. H. H. (2020). Reflection on modern methods: Five myths about measurement error in epidemiological research. International Journal of Epidemiology, 49(1), 338–347. https://doi.org/10.1093/ije/dyz251
-
-
www.thelancet.com www.thelancet.com
-
Lozano, R., Fullman, N., Mumford, J. E., Knight, M., Barthelemy, C. M., Abbafati, C., Abbastabar, H., Abd-Allah, F., Abdollahi, M., Abedi, A., Abolhassani, H., Abosetugn, A. E., Abreu, L. G., Abrigo, M. R. M., Haimed, A. K. A., Abushouk, A. I., Adabi, M., Adebayo, O. M., Adekanmbi, V., … Murray, C. J. L. (2020). Measuring universal health coverage based on an index of effective coverage of health services in 204 countries and territories, 1990–2019: A systematic analysis for the Global Burden of Disease Study 2019. The Lancet, 0(0). https://doi.org/10.1016/S0140-6736(20)30750-9
-
-
onlinelibrary.wiley.com onlinelibrary.wiley.com
-
Collins, G. S., & Wilkinson, J. (n.d.). Statistical issues in the development a COVID-19 prediction models. Journal of Medical Virology, n/a(n/a). https://doi.org/10.1002/jmv.26390
-
-
covid-19.iza.org covid-19.iza.org
-
On the Effects of COVID-19 Safer-At-Home Policies on Social Distancing, Car Crashes and Pollution. (n.d.). IZA – Institute of Labor Economics. Retrieved August 4, 2020, from https://covid-19.iza.org/publications/dp13255/
-
-
covid-19.iza.org covid-19.iza.org
-
Public Attention and Policy Responses to COVID-19 Pandemic. COVID-19 and the Labor Market. (n.d.). IZA – Institute of Labor Economics. Retrieved July 31, 2020, from https://covid-19.iza.org/publications/dp13427/
-
-
covid-19.iza.org covid-19.iza.org
-
Does the COVID-19 Pandemic Improve Global Air Quality? New Cross-National Evidence on Its Unintended Consequences. COVID-19 and the Labor Market. (n.d.). IZA – Institute of Labor Economics. Retrieved July 29, 2020, from https://covid-19.iza.org/publications/dp13480/
-
- Jul 2020
-
gss.civilservice.gov.uk gss.civilservice.gov.uk
-
An analyst’s job is never done – GSS. (n.d.). Retrieved July 3, 2020, from https://gss.civilservice.gov.uk/blog/an-analysts-job-is-never-done/
-
- Jun 2020
-
psyarxiv.com psyarxiv.com
-
Antonakis, J., Bastardoz, N., & Jacquart, P. (2020). In praise of the impact factor [Preprint]. PsyArXiv. https://doi.org/10.31234/osf.io/h4p9e
-
- Apr 2020
-
www.nationalreview.com www.nationalreview.com
-
Verbruggen, R. (2020 March 24). Another COVID Cost-Benefit Analysis. National Review. https://www.nationalreview.com/corner/another-covid-cost-benefit-analysis/
-
- Aug 2018
-
www.oxfam.org.uk www.oxfam.org.uk
-
13
13
-
4
4
-
2
2
-
- May 2018
-
www.audit.vic.gov.au www.audit.vic.gov.au
-
Negative values included when assessing air quality In computing average pollutant concentrations, EPA includes recorded values that are below zero. EPA advised that this is consistent with NEPM AAQ procedures. Logically, however, the lowest possible value for air pollutant concentrations is zero. Either it is present, even if in very small amounts, or it is not. Negative values are an artefact of the measurement and recording process. Leaving negative values in the data introduces a negative bias, which potentially under represents actual concentrations of pollutants. We noted a considerable number of negative values recorded. For example, in 2016, negative values comprised 5.3 per cent of recorded hourly PM2.5 values, and 1.3 per cent of hourly PM10 values. When we excluded negative values from the calculation of one‐day averages, there were five more exceedance days for PM2.5 and one more for PM10 during 2016.
-
- May 2017
-
www.safaribooksonline.com www.safaribooksonline.com
-
volume, velocity, and variety
volume: The actual size of traffic
Velocity: How fast does the traffic show up.
Variety: Refers to data that can be unstructured, semi structured or multi structured.
-