Based on the Pathfinder sea surface temperature(PFSST),the surface axis and its pattern of the Yellow Sea Warm Current(YSWC) are discussed.A structure of double-warm-tongue is found in February and it varies in differ...Based on the Pathfinder sea surface temperature(PFSST),the surface axis and its pattern of the Yellow Sea Warm Current(YSWC) are discussed.A structure of double-warm-tongue is found in February and it varies in different years.Two indexes are calculated to represent the westward shift(WSI) and northward extension(NEI) of the warm water in the Yellow Sea(YS).Wavelet analysis illustrates that the WSI and NEI have prominent periods of 3-6 years and 3-4 years,respectively.The Empirical Orthogonal Function(EOF) decomposition is applied to the winter wind stress curl and the Kuroshio Current(KC) transport,which are believed to play important roles in forcing the variability of the YSWC surface axis.Statistics shows that the WSI is significantly related with the second EOF mode of the wind stress curl in February,which may force the YSWC surface axis moving westward and maintaining the double warm tongues because of its opposite curl in the YSWC domain.The first EOF mode of wind stress curl in January is propitious for inducing the warm tongue in the YS to advance more northward.Hence,the wind stress curls both in January and in February could force variations of the YSWC surface axis;however,the effect of the January wind stress curl is relatively weaker than that of the February.The relationship between the NEI and the KC transport is remarkable,and it seems that the stronger KC supplies more power to push the YSWC northward against the southward wind.展开更多
The South China Sea warm pool interacts vigorously with the summer monsoon which is active in the region. However, there has not been a definition concerning the former warm pool which is as specific as that for the l...The South China Sea warm pool interacts vigorously with the summer monsoon which is active in the region. However, there has not been a definition concerning the former warm pool which is as specific as that for the latter. The seasonal and inter-annual variability of the South China Sea warm pool and its relations to the South China Sea monsoon onset were analyzed using Levitus and NCEP/NCAR OISST data. The results show that, the seasonal variability of the South China Sea warm pool is obvious, which is weak in winter, develops rapidly in spring, becomes strong and extensive in summer and early autumn, and quickly decays from mid-autumn. The South China Sea warm pool is 55 m in thickness in the strongest period and its axis is oriented from southwest to northeast with the main section locating along the western offshore steep slope of northern Kalimantan-Palawan Island. For the warm pools in the South China Sea, west Pacific and Indian Ocean, the oscillation, which is within the same large scale air-sea coupling system, is periodic around 5 years. There are additional oscillations of about 2.5 years and simultaneous inter-annual variations for the latter two warm pools. The intensity of the South China Sea warm pool varies by a lag of about 5 months as compared to the west Pacific one. The result also indicates that the inter-annual variation of the intensity index is closely related with the onset time of the South China Sea monsoon. When the former is persistently warmer (colder) in preceding winter and spring, the monsoon in the South China Sea usually sets in on a later (earlier) date in early summer. The relation is associated with the activity of the high pressure over the sea in early summer. An oceanic background is given for the prediction of the South China Sea summer monsoon, though the mechanism through which the warm pool and eventually the monsoon are affected remains unclear.展开更多
The inter-annual variability of rainfall onset and crop replanting in East Africa (EA) was assessed using daily estimated rainfall data from climate hazard group infrared precipitation (CHIRPS Ver2.0) and monthly Sea ...The inter-annual variability of rainfall onset and crop replanting in East Africa (EA) was assessed using daily estimated rainfall data from climate hazard group infrared precipitation (CHIRPS Ver2.0) and monthly Sea Surface Temperature (SST) indices [Indian Ocean Dipole (IOD) and El-Ni?o Southern Oscillation (ENSO) at NINO3.4 region] from the National Center for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR). The data covered a period of 40 years from1981 to 2020. The methods of cumulative of daily mean rainfall, percentage of onset date departure (PODD), Mann-Kendall (MK) trend test, student t-test, and correlation were applied in the analysis. The results showed that early onset with dry spell (WDS) consideration frequently occurs in Uganda between the first and second dekads of September, while late rainfall onset WDS occurs in the first and second dekads of December over central and Northern Kenya as well as in the Northeastern highlands, parts of the northern coast and unimodal regions in Tanzania. Rainfall onset with no dry spell (WnDS) portrayed an average of 10 days before the occurrence of true onset WDS, with maximum onset departure days (ODD) above 30 days across the Rift Valley area in Kenya and the Northeastern highlands in Tanzania. The high chance of minimum ODD is seen over entire Uganda and the area around Lake Victoria. However, few regions, such as Nakuru (Kenya) Gulu and Kibale (Uganda), and Gitega (Burundi), revealed a slight positive linear trend while others showed negative trend. Significant positive patterns for correlation between onset WDS and SST indices (IOD and NINO 3.4) were discovered in Northern and Northeastern Kenya, as well as areas along the Indian Ocean (over Tanzania’s Northern Coast). Inter-annual relationship between onset dates WDS and IOD (NINO3.4) indices exhibits a high correlation coefficient r = 0.23 (r = 0.48) in Uganda and r = 0.44 (r = 0.36) in Kenya. On the other hand, a negative correlation was revealed over Burundi and Tanzania (over a unimodal region). A high percentage of PODD was observed, ranging from 40% to 70% over the Rift Valley in Kenya and at the Northeastern highlands in Tanzania. However, a strong PODD above 70% was observed over Tanga and the Northern Pwani Region in Tanzania. These findings will help farmers to understand the appropriate time for crop planting, as well as help other socio-economic activities that strongly depend on rainfall.展开更多
A time series dataset spanning 39 years(1981-2018) on red tide events in Zhejiang coastal waters was used to study the characteristics of inter-annual spatial and temporal variations. A distinct inter-annual pattern c...A time series dataset spanning 39 years(1981-2018) on red tide events in Zhejiang coastal waters was used to study the characteristics of inter-annual spatial and temporal variations. A distinct inter-annual pattern characterized by low frequency, explosive growth and fluctuating decline stages was found over the studied time scale. Most red tide events occurred in parallel to the bathymetric contour, and 95.4% were located to the west of the 50 m isobath. Additionally, the high-incidence area of red tides is expanding southward. In this paper, local sea surface temperature(SST), mariculture area and secondary industry growth rate are introduced and identified as the main factors influencing the nutrient and hydrometeorological conditions. A multivariate nonlinear regression equation based on these factors was constructed, and the goodness of fit coefficient was 0.907. The causes of the annual variation and high-frequency area in the southward expansion were quantitatively analyzed based on the proposed regression model. Finally, the results indicated that 68.7% of the annual occurrence variation of red tide was due to the SST and mariculture area, which are the main impact factors;however,secondary industry growth could compensate for the nutrient deficiency caused by the sharp mariculture area reduction and decreased SST. The background nutrient level, which is elevated by coastal economic development, especially secondary industry, is the main determinant of the southward expansion. Although the trend of the southward expansion of high-frequency areas has not changed, the red tide frequency in coastal cities has decreased by half and remained at a stable level after 2010 due to substantial economic restructuring and environmental protection.展开更多
Based on rotational empirical orthogonal function(REOF),max-entropy and Mexico-hat wavelet transform techniques,monthly SSTA of the tropical Pacific,Atlantic and Indian Oceans (32.5°S—32. 5°N)is investigate...Based on rotational empirical orthogonal function(REOF),max-entropy and Mexico-hat wavelet transform techniques,monthly SSTA of the tropical Pacific,Atlantic and Indian Oceans (32.5°S—32. 5°N)is investigated.It is shown that the inter-decadal variability and inter-annual variability take on global scale,and there exist their own significance areas.Moreover,through the total time series,the intensity of the variabilities is time-variable.And in fact,both the variabilities are usually coexistent.In significance areas of each of the variabilities,another variability is sometimes quite strong.展开更多
Climate variability has become a decisive factor for the crop yields in Sub-Saharan African countries. Inter-annual variability in rainfall has been the key climatic element that determines the success of agriculture ...Climate variability has become a decisive factor for the crop yields in Sub-Saharan African countries. Inter-annual variability in rainfall has been the key climatic element that determines the success of agriculture in this region.From an analysis of recent rainfall conditions in West Africa,FAO(2001) concluded that a long-term change in rainfall has occurred in the semi-arid and sub-humid zones of West Africa.Thus,this study aims at assessing the vulnerability of crop yields to inter-annual variability in rainfall in sub-Saharan African countries taking Nigeria as a case study.A time series of the averaged value of the 12-Month Weighted Anomaly Standardized Precipita-展开更多
Indian Summer Monsoon Rainfall （ISMR） exhibits a prominent inter-annual variability known as troposphere biennial oscillation.A season of deficient June to September monsoon rainfall in India is followed by warm sea...Indian Summer Monsoon Rainfall （ISMR） exhibits a prominent inter-annual variability known as troposphere biennial oscillation.A season of deficient June to September monsoon rainfall in India is followed by warm sea surface temperature （SST） anomalies over the tropical Indian Ocean and cold SST anomalies over the westem Pacific Ocean.These anomalies persist until the following monsoon,which yields normal or excessive rainfall.Monsoon rainfall in India has shown decadal variability in the form of 30 year epochs of alternately occurring frequent and infrequent drought monsoons since 1841,when rainfall measurements began in India.Decadal oscillations of monsoon rainfall and the well known decadal oscillations in SSTs of the Atlantic and Pacific oceans have the same period of approximately 60 years and nearly the same temporal phase.In both of these variabilities,anomalies in monsoon heat source,such as deep convection,and middle latitude westerlies of the upper troposphere over south Asia have prominent roles.展开更多
In this paper,we establish some strong laws of large numbers,which are for nonindependent random variables under the framework of sublinear expectations.One of our main results is for blockwise m-dependent random vari...In this paper,we establish some strong laws of large numbers,which are for nonindependent random variables under the framework of sublinear expectations.One of our main results is for blockwise m-dependent random variables,and another is for sub-orthogonal random variables.Both extend the strong law of large numbers for independent random variables under sublinear expectations to the non-independent case.展开更多
Inter-annual variability of the Kuroshio water intrusion on the shelf of East China Sea (ECS) was simulated with a nested global and Northwest Pacific ocean circulation model.The model analysis reveals the influence o...Inter-annual variability of the Kuroshio water intrusion on the shelf of East China Sea (ECS) was simulated with a nested global and Northwest Pacific ocean circulation model.The model analysis reveals the influence of the variability of Kuroshio transport east of Taiwan on the intrusion to the northeast of Taiwan:high correlation (r=0.92) with the on-shore volume flux in the lower layer (50 200 m) ;low correlation (r=0.50) with the on-shore flux in the upper layer (0 50 m) .Spatial distribution of correlations between volume fluxes and sea surface height suggests that inter-annual variability of the Kuroshio flux east of Taiwan and its subsurface water intruding to the shelf lag behind the sea surface height anomalies in the central Pacific at 162 E by about 14 months,and could be related to wind-forced variation in the interior North Pacific that propagates westward as Rossby waves.The intrusion of Kuroshio surface water is also influenced by local winds.The intruding Kuroshio subsurface water causes variations of temperature and salinity of bottom waters on the southern ECS shelf.The influence of the intruding Kuroshio subsurface water extends widely from the shelf slope northeast of Taiwan northward to the central ECS near the 60 m isobath,and northeastward to the region near the 90 m isobath.展开更多
The aim of the paper is to analyze a possible teleconnection of Quasi-Biennial Oscillation （QBO）, Southern Oscillation （SO）, North Atlantic Oscillation （NAO）, and Arctic Oscillation （AO） phenomena with longter...The aim of the paper is to analyze a possible teleconnection of Quasi-Biennial Oscillation （QBO）, Southern Oscillation （SO）, North Atlantic Oscillation （NAO）, and Arctic Oscillation （AO） phenomena with longterm streamflow fluctuation of the Bela River （1895-2004） and Cierny Hron River （1931-2004） （central Slovakia）. Homogeneity, long-term trends, as well as inter-annual dry and wet cycles were analyzed for the entire 1895-2004 time series of the Bela River and for the 1931-2004 time series of the Cierny Hron River. Inter-annual fluctuation of the wet and dry periods was identified using spectral analysis. The most significant period is that of 3.6 years. Other significant periods are those of 2.35 years, 13.5 years, and 21 years. Since these periods were found in other rivers of the world, as well as in SO, NAO, and AO phenomena, they can be considered as relating to the general regularity of the Earth.展开更多
The study on how the variations in CO2 sources and sinks can affect the CO2 concentration over East Asia would be useful to provide information for policymaker concerning carbon emission reduction.In this study,a nest...The study on how the variations in CO2 sources and sinks can affect the CO2 concentration over East Asia would be useful to provide information for policymaker concerning carbon emission reduction.In this study,a nested-grid version of global chemical transport model(GEOS-Chem)is employed to assess the impacts of variations in meteorological parameters,terrestrial fluxes,fossil fuel emissions,and biomass burning on inter-annual variations of CO2 concentrations over East Asia in 2004—2012.Simulated CO2 concentrations are compared with observations at 14 surface stations from the World Data Centre for Greenhouse Gases(WDCGG)and satellite-derived C 02 column density(XCO,)from the Gases Observing SATellite(GOSAT).The comparison shows that the simulated CO2 column density is generally higher than that of GOSAT by 1.33×10^6(annual mean point by point biases averaged over East Asia).The model reasonably captures the temporal variations of CO2 concentrations observed at the ground-based stations,but it is likely to underestimate the peaks-to-troughs amplitude of the seasonal cycle by 50%or more.The simulated surface CO2 concentration in East Asia exhibits the largest inter-annual variation in December-January—February(DJF).The regional mean absolute deviation(MAD)values over East Asia are within(4.4—5.0)×10^-6 for all seasons.Model sensitivity simulations indicate that the inter-annual variations of surface CO2 concentrations are mainly driven by variations of meteorological parameters,and partly modulated by the inter-annual variations of terrestrial fluxes and fossil fuel emissions in local regions.The variations of the terrestrial fluxes and fossil fuel emissions may account for〜28%of the inter-annual variation of surface CO2 concentration in southern China.The inter-annual variations of the peaks-to-troughs amplitude are dependent on variations of meteorological parameters,terrestrial fluxes and fossil fuel emissions in local regions.The influence of biomass burning emissions is relatively weak.展开更多
On the basis of the CTD data obtained within the Bering Sea shelf by the Second to Sixth Chinese National Arctic Research Expedition in the summers of 2003, 2008, 2010, 2012 and 2014, the classification and interannua...On the basis of the CTD data obtained within the Bering Sea shelf by the Second to Sixth Chinese National Arctic Research Expedition in the summers of 2003, 2008, 2010, 2012 and 2014, the classification and interannual variation of water masses on the central Bering Sea shelf and the northern Bering Sea shelf are analyzed. The results indicate that there are both connection and difference between two regions in hydrological features. On the central Bering Sea shelf, there are mainly four types of water masses distribute orderly from the slope to the coast of Alaska： Bering Slope Current Water（BSCW）, MW（Mixed Water）, Bering Shelf Water（BSW） and Alaska Coastal Water（ACW）. In summer, BSW can be divided into Bering Shelf Surface Water（BSW_S） and Bering Shelf Cold Water（BSW_C）. On the northern Bering Sea shelf near the Bering Strait,it contains Anadyr Water（AW）, BSW and ACW from west to east. But the spatial-temporal features are also remarkable in each region. On the central shelf, the BSCW is saltiest and occupies the west of 177°W, which has the highest salinity in 2014. The BSW_C is the coldest water mass and warmest in 2014; the ACW is freshest and mainly occupies the east of 170°W, which has the highest temperature and salinity in 2012. On the northern Bering Sea shelf near the Bering Strait, the AW is saltiest with temperature decreasing sharply compared with BSCW on the central shelf. In the process of moving northward to the Bering Strait, the AW demonstrates a trend of eastward expansion. The ACW is freshest but saltier than the ACW on the central shelf,which is usually located above the BSW and is saltiest in 2014. The BSW distributes between the AW and the ACW and coldest in 2012, but the cold water of the BSW_C on the central shelf, whose temperature less than 0°C, does not exist on the northern shelf. Although there are so many changes, the respond to a climate change is synchronized in the both regions, which can be divided into the warm years（2003 and 2014） and cold years（2008, 2010 and 2012）. The year of 2014 may be a new beginning of warm period.展开更多
Using more than 14 years of GRACE(Gravity Recovery and Climate Experiment) satellite gravimetry observations, we estimate the ice loss rate for the Patagonia Ice Field(PIF) of South America. After correcting the effec...Using more than 14 years of GRACE(Gravity Recovery and Climate Experiment) satellite gravimetry observations, we estimate the ice loss rate for the Patagonia Ice Field(PIF) of South America. After correcting the effects of glacier isostatic adjustment(GIA) and hydrological variations, the ice loss rate is -23.5 ± 8.1 Giga ton per year(Gt/yr) during the period April 2002 through December 2016, equivalent to an average ice thickness change of-1.3 m/yr if evenly distributed over PIF. The PIF ice mass change series also show obvious inter-annual variations during the entire period. For the time spans April 2002 to December 2007, January 2008 to December 2012 and January 2013 to December 2016, the ice loss rates are -26.4,-9.0 and -25.0 Gt/yr, respectively, indicating that the ice melting experienced significant slowing down and accelerating again in the past decade. Comparison with time series from temperature and precipitation data over PIF suggests that the inter-annual ice losses might not be directly correlated with the temperature changes and precipitation anomalies, and thus their interrelation is intricate. However, the dramatic ice loss acceleration in 2016(with more than 100 Gt within the first half of the year) appears closely related with the evident temperature increase and severe precipitation shortage over 2016, which are likely correlated with the strong E1 Nino event around 2016. Moreover, we compare the GRACE spherical harmonic(SH) and mass concentration(Mascon) solutions in estimating the PIF ice loss rate, and find that the Mascon result has larger uncertainty in leakage error correction,while the SH solutions can better correct leakage errors based on a constrained forward modeling iterative method. Thus the GRACE SH solutions with constrained forward modeling recovery are recommended to evaluating the ice mass change of PIF or other glacier regions with relatively smaller spatial scales.展开更多
Plants play an essential role in matter and energy transformations and are key messengers in the carbon and energy cycle. Net primary productivity (NPP) reflects the capability of plants to transform solar energy into...Plants play an essential role in matter and energy transformations and are key messengers in the carbon and energy cycle. Net primary productivity (NPP) reflects the capability of plants to transform solar energy into photosynthesis. It is very sensible for factors affecting on vegetation variability such as climate, soils, plant characteristics and human activities. So, it can be used as an indicator of actual and potential trend of vegetation. In this study we used the actual NPP which was derived from MODIS to assess the response of NPP to climate variables in Gadarif State, from 2000 to 2010. The correlations between NPP and climate variables (temperature and precipitation) are calculated using Pearson’s Correlation Coefficient and ordinary least squares regression. The main results show the following 1) the correlation Coefficient between NPP and mean annual temperature is Somewhat negative for Feshaga, Rahd, Gadarif and Galabat areas and weakly negative in Faw area;2) the correlation Coefficient between NPP and annual total precipitation is weakly negative in Faw, Rahd and Galabat areas and somewhat negative in Galabat and Rahd areas. This study demonstrated that the correlation analysis between NPP and climate variables (precipitation and temperature) gives reliably result of NPP responses to climate variables that is clearly in a very large scale of study area.展开更多
An internal state variable(ISV)model was established according to the experimental results of hot plane strain compression(PSC)to predict the microstructure evolution during hot spinning of ZK61 alloy.The effects of t...An internal state variable(ISV)model was established according to the experimental results of hot plane strain compression(PSC)to predict the microstructure evolution during hot spinning of ZK61 alloy.The effects of the internal variables were considered in this ISV model,and the parameters were optimized by genetic algorithm.After validation,the ISV model was used to simulate the evolution of grain size(GS)and dynamic recrystallization(DRX)fraction during hot spinning via Abaqus and its subroutine Vumat.By comparing the simulated results with the experimental results,the application of the ISV model was proven to be reliable.Meanwhile,the strength of the thin-walled spun ZK61 tube increased from 303 to 334 MPa due to grain refinement by DRX and texture strengthening.Besides,some ultrafine grains(0.5μm)that played an important role in mechanical properties were formed due to the proliferation,movement,and entanglement of dislocations during the spinning process.展开更多
In order to solve the problem of the variable coefficient ordinary differen-tial equation on the bounded domain,the Lagrange interpolation method is used to approximate the exact solution of the equation,and the error...In order to solve the problem of the variable coefficient ordinary differen-tial equation on the bounded domain,the Lagrange interpolation method is used to approximate the exact solution of the equation,and the error between the numerical solution and the exact solution is obtained,and then compared with the error formed by the difference method,it is concluded that the Lagrange interpolation method is more effective in solving the variable coefficient ordinary differential equation.展开更多
As a crucial component of the Earth’s climate system,Antarctic sea ice has demonstrated significant variability over the satellite era.Here,we identify a remarkable decadal transition in the total Antarctic Sea Ice E...As a crucial component of the Earth’s climate system,Antarctic sea ice has demonstrated significant variability over the satellite era.Here,we identify a remarkable decadal transition in the total Antarctic Sea Ice Extent(SIE).The stage from 1979 to 2006 is characterized by high-frequency(i.e.,seasonal to interannual)temporal variability in SIE and zonal asymmetry in Sea Ice Concentration(SIC),which is primarily under the control of the Amundsen Sea Low(ASL).After 2007,however,sea ice changes exhibit a more spatially homogeneous pattern in SIC and a more temporally long-lasting mode in SIE.Further analysis reveals that sea ice-ocean interaction plays a major role in the low-frequency(i.e.,multiannual)variability of Antarctic sea ice from 2007−22.The related physical process is inferred to manifest as a strong coupling between the surface and the subsurface ocean layers,involving enhanced vertical convection and the downward delivery of the surface anomalies related to ice melting and freezing processes,thus maintaining the SIE anomalies for a longer time.Furthermore,this process mainly occurs in the Amundsen-Bellingshausen Sea(ABS)sector,and the weakened subsurface ocean stratification is the key factor triggering the coupling process in this region.We find that the Circumpolar Deep Water(CDW)over the ABS sector continued to shoal before 2007 and remained stable thereafter.It is speculated that the shoaling of the CDW may be a possible driver leading to the weakening of the subsurface stratification.展开更多
El Niño-Southern Oscillation(ENSO)is a major driver of climate change in middle and low latitudes and thus strongly influences the terrestrial carbon cycle through land-air interaction.Both the ENSO modulation an...El Niño-Southern Oscillation(ENSO)is a major driver of climate change in middle and low latitudes and thus strongly influences the terrestrial carbon cycle through land-air interaction.Both the ENSO modulation and carbon flux variability are projected to increase in the future,but their connection still needs further investigation.To investigate the impact of future ENSO modulation on carbon flux variability,this study used 10 CMIP6 earth system models to analyze ENSO modulation and carbon flux variability in middle and low latitudes,and their relationship,under different scenarios simulated by CMIP6 models.The results show a high consistency in the simulations,with both ENSO modulation and carbon flux variability showing an increasing trend in the future.The higher the emissions scenario,especially SSP5-8.5 compared to SSP2-4.5,the greater the increase in variability.Carbon flux variability in the middle and low latitudes under SSP2-4.5 increases by 30.9%compared to historical levels during 1951-2000,while under SSP5-8.5 it increases by 58.2%.Further analysis suggests that ENSO influences mid-and low-latitude carbon flux variability primarily through temperature.This occurrence may potentially be attributed to the increased responsiveness of gross primary productivity towards regional temperature fluctuations,combined with the intensified influence of ENSO on land surface temperatures.展开更多
Large-scale Language Models(LLMs)have achieved significant breakthroughs in Natural Language Processing(NLP),driven by the pre-training and fine-tuning paradigm.While this approach allows models to specialize in speci...Large-scale Language Models(LLMs)have achieved significant breakthroughs in Natural Language Processing(NLP),driven by the pre-training and fine-tuning paradigm.While this approach allows models to specialize in specific tasks with reduced training costs,the substantial memory requirements during fine-tuning present a barrier to broader deployment.Parameter-Efficient Fine-Tuning(PEFT)techniques,such as Low-Rank Adaptation(LoRA),and parameter quantization methods have emerged as solutions to address these challenges by optimizing memory usage and computational efficiency.Among these,QLoRA,which combines PEFT and quantization,has demonstrated notable success in reducing memory footprints during fine-tuning,prompting the development of various QLoRA variants.Despite these advancements,the quantitative impact of key variables on the fine-tuning performance of quantized LLMs remains underexplored.This study presents a comprehensive analysis of these key variables,focusing on their influence across different layer types and depths within LLM architectures.Our investigation uncovers several critical findings:(1)Larger layers,such as MLP layers,can maintain performance despite reductions in adapter rank,while smaller layers,like self-attention layers,aremore sensitive to such changes;(2)The effectiveness of balancing factors depends more on specific values rather than layer type or depth;(3)In quantization-aware fine-tuning,larger layers can effectively utilize smaller adapters,whereas smaller layers struggle to do so.These insights suggest that layer type is a more significant determinant of fine-tuning success than layer depth when optimizing quantized LLMs.Moreover,for the same discount of trainable parameters,reducing the trainable parameters in a larger layer is more effective in preserving fine-tuning accuracy than in a smaller one.This study provides valuable guidance for more efficient fine-tuning strategies and opens avenues for further research into optimizing LLM fine-tuning in resource-constrained environments.展开更多
During the boreal summer,intraseasonal oscillations exhibit significant interannual variations in intensity over two key regions:the central-western equatorial Pacific(5°S-5°N,150°E-150°W)and the s...During the boreal summer,intraseasonal oscillations exhibit significant interannual variations in intensity over two key regions:the central-western equatorial Pacific(5°S-5°N,150°E-150°W)and the subtropical Northwestern Pacific(10°-20°N,130°E-175°W).The former is well-documented and considered to be influenced by the ENSO,while the latter has received comparatively less attention and is likely influenced by the Pacific Meridional Mode(PMM),as suggested by partial correlation analysis results.To elucidate the physical processes responsible for the enhanced(weakened)intraseasonal convection over the subtropical northwestern Pacific during warm(cold)PMM years,the authors employed a moisture budget analysis.The findings reveal that during warm PMM years,there is an increase in summer-mean moisture over the subtropical northwestern Pacific.This increase interacts with intensified vertical motion perturbations in the region,leading to greater vertical moisture advection in the lower troposphere and consequently resulting in convective instability.Such a process is pivotal in amplifying intraseasonal convection anomalies.The observational findings were further verified by model experiments forced by PMM-like sea surface temperature patterns.展开更多
基金Supported by the National Basic Research Program of China (973 Program) (No 2005C B422308)the National High-tech Research and Development Program (863 Program) (No 2006AA09Z149)the China International Science and Technology Cooperation Program (No2006DFB21250)
文摘Based on the Pathfinder sea surface temperature(PFSST),the surface axis and its pattern of the Yellow Sea Warm Current(YSWC) are discussed.A structure of double-warm-tongue is found in February and it varies in different years.Two indexes are calculated to represent the westward shift(WSI) and northward extension(NEI) of the warm water in the Yellow Sea(YS).Wavelet analysis illustrates that the WSI and NEI have prominent periods of 3-6 years and 3-4 years,respectively.The Empirical Orthogonal Function(EOF) decomposition is applied to the winter wind stress curl and the Kuroshio Current(KC) transport,which are believed to play important roles in forcing the variability of the YSWC surface axis.Statistics shows that the WSI is significantly related with the second EOF mode of the wind stress curl in February,which may force the YSWC surface axis moving westward and maintaining the double warm tongues because of its opposite curl in the YSWC domain.The first EOF mode of wind stress curl in January is propitious for inducing the warm tongue in the YS to advance more northward.Hence,the wind stress curls both in January and in February could force variations of the YSWC surface axis;however,the effect of the January wind stress curl is relatively weaker than that of the February.The relationship between the NEI and the KC transport is remarkable,and it seems that the stronger KC supplies more power to push the YSWC northward against the southward wind.
基金A comprehensive study on the activities of the South China Sea summer monsoon and its influence"- a major project of the Chines
文摘The South China Sea warm pool interacts vigorously with the summer monsoon which is active in the region. However, there has not been a definition concerning the former warm pool which is as specific as that for the latter. The seasonal and inter-annual variability of the South China Sea warm pool and its relations to the South China Sea monsoon onset were analyzed using Levitus and NCEP/NCAR OISST data. The results show that, the seasonal variability of the South China Sea warm pool is obvious, which is weak in winter, develops rapidly in spring, becomes strong and extensive in summer and early autumn, and quickly decays from mid-autumn. The South China Sea warm pool is 55 m in thickness in the strongest period and its axis is oriented from southwest to northeast with the main section locating along the western offshore steep slope of northern Kalimantan-Palawan Island. For the warm pools in the South China Sea, west Pacific and Indian Ocean, the oscillation, which is within the same large scale air-sea coupling system, is periodic around 5 years. There are additional oscillations of about 2.5 years and simultaneous inter-annual variations for the latter two warm pools. The intensity of the South China Sea warm pool varies by a lag of about 5 months as compared to the west Pacific one. The result also indicates that the inter-annual variation of the intensity index is closely related with the onset time of the South China Sea monsoon. When the former is persistently warmer (colder) in preceding winter and spring, the monsoon in the South China Sea usually sets in on a later (earlier) date in early summer. The relation is associated with the activity of the high pressure over the sea in early summer. An oceanic background is given for the prediction of the South China Sea summer monsoon, though the mechanism through which the warm pool and eventually the monsoon are affected remains unclear.
文摘The inter-annual variability of rainfall onset and crop replanting in East Africa (EA) was assessed using daily estimated rainfall data from climate hazard group infrared precipitation (CHIRPS Ver2.0) and monthly Sea Surface Temperature (SST) indices [Indian Ocean Dipole (IOD) and El-Ni?o Southern Oscillation (ENSO) at NINO3.4 region] from the National Center for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR). The data covered a period of 40 years from1981 to 2020. The methods of cumulative of daily mean rainfall, percentage of onset date departure (PODD), Mann-Kendall (MK) trend test, student t-test, and correlation were applied in the analysis. The results showed that early onset with dry spell (WDS) consideration frequently occurs in Uganda between the first and second dekads of September, while late rainfall onset WDS occurs in the first and second dekads of December over central and Northern Kenya as well as in the Northeastern highlands, parts of the northern coast and unimodal regions in Tanzania. Rainfall onset with no dry spell (WnDS) portrayed an average of 10 days before the occurrence of true onset WDS, with maximum onset departure days (ODD) above 30 days across the Rift Valley area in Kenya and the Northeastern highlands in Tanzania. The high chance of minimum ODD is seen over entire Uganda and the area around Lake Victoria. However, few regions, such as Nakuru (Kenya) Gulu and Kibale (Uganda), and Gitega (Burundi), revealed a slight positive linear trend while others showed negative trend. Significant positive patterns for correlation between onset WDS and SST indices (IOD and NINO 3.4) were discovered in Northern and Northeastern Kenya, as well as areas along the Indian Ocean (over Tanzania’s Northern Coast). Inter-annual relationship between onset dates WDS and IOD (NINO3.4) indices exhibits a high correlation coefficient r = 0.23 (r = 0.48) in Uganda and r = 0.44 (r = 0.36) in Kenya. On the other hand, a negative correlation was revealed over Burundi and Tanzania (over a unimodal region). A high percentage of PODD was observed, ranging from 40% to 70% over the Rift Valley in Kenya and at the Northeastern highlands in Tanzania. However, a strong PODD above 70% was observed over Tanga and the Northern Pwani Region in Tanzania. These findings will help farmers to understand the appropriate time for crop planting, as well as help other socio-economic activities that strongly depend on rainfall.
基金The National Basic Research Program of China under contract No. 2016YFC1401900the Open Research Funds of Key Laboratory of Marine Ecological Monitoring and Restoration Technologies under contract No. MATHAB201703。
文摘A time series dataset spanning 39 years(1981-2018) on red tide events in Zhejiang coastal waters was used to study the characteristics of inter-annual spatial and temporal variations. A distinct inter-annual pattern characterized by low frequency, explosive growth and fluctuating decline stages was found over the studied time scale. Most red tide events occurred in parallel to the bathymetric contour, and 95.4% were located to the west of the 50 m isobath. Additionally, the high-incidence area of red tides is expanding southward. In this paper, local sea surface temperature(SST), mariculture area and secondary industry growth rate are introduced and identified as the main factors influencing the nutrient and hydrometeorological conditions. A multivariate nonlinear regression equation based on these factors was constructed, and the goodness of fit coefficient was 0.907. The causes of the annual variation and high-frequency area in the southward expansion were quantitatively analyzed based on the proposed regression model. Finally, the results indicated that 68.7% of the annual occurrence variation of red tide was due to the SST and mariculture area, which are the main impact factors;however,secondary industry growth could compensate for the nutrient deficiency caused by the sharp mariculture area reduction and decreased SST. The background nutrient level, which is elevated by coastal economic development, especially secondary industry, is the main determinant of the southward expansion. Although the trend of the southward expansion of high-frequency areas has not changed, the red tide frequency in coastal cities has decreased by half and remained at a stable level after 2010 due to substantial economic restructuring and environmental protection.
基金National Key Project-Studies on Short-Term Climate Prediction Systems in China
文摘Based on rotational empirical orthogonal function(REOF),max-entropy and Mexico-hat wavelet transform techniques,monthly SSTA of the tropical Pacific,Atlantic and Indian Oceans (32.5°S—32. 5°N)is investigated.It is shown that the inter-decadal variability and inter-annual variability take on global scale,and there exist their own significance areas.Moreover,through the total time series,the intensity of the variabilities is time-variable.And in fact,both the variabilities are usually coexistent.In significance areas of each of the variabilities,another variability is sometimes quite strong.
文摘Climate variability has become a decisive factor for the crop yields in Sub-Saharan African countries. Inter-annual variability in rainfall has been the key climatic element that determines the success of agriculture in this region.From an analysis of recent rainfall conditions in West Africa,FAO(2001) concluded that a long-term change in rainfall has occurred in the semi-arid and sub-humid zones of West Africa.Thus,this study aims at assessing the vulnerability of crop yields to inter-annual variability in rainfall in sub-Saharan African countries taking Nigeria as a case study.A time series of the averaged value of the 12-Month Weighted Anomaly Standardized Precipita-
基金the European Commission(Project INDO-MARECLIM)the Norwegian Research Council(Project INDIA-CLIM)for providing financial support for this study
文摘Indian Summer Monsoon Rainfall （ISMR） exhibits a prominent inter-annual variability known as troposphere biennial oscillation.A season of deficient June to September monsoon rainfall in India is followed by warm sea surface temperature （SST） anomalies over the tropical Indian Ocean and cold SST anomalies over the westem Pacific Ocean.These anomalies persist until the following monsoon,which yields normal or excessive rainfall.Monsoon rainfall in India has shown decadal variability in the form of 30 year epochs of alternately occurring frequent and infrequent drought monsoons since 1841,when rainfall measurements began in India.Decadal oscillations of monsoon rainfall and the well known decadal oscillations in SSTs of the Atlantic and Pacific oceans have the same period of approximately 60 years and nearly the same temporal phase.In both of these variabilities,anomalies in monsoon heat source,such as deep convection,and middle latitude westerlies of the upper troposphere over south Asia have prominent roles.
文摘In this paper,we establish some strong laws of large numbers,which are for nonindependent random variables under the framework of sublinear expectations.One of our main results is for blockwise m-dependent random variables,and another is for sub-orthogonal random variables.Both extend the strong law of large numbers for independent random variables under sublinear expectations to the non-independent case.
基金supported by the National Basic Research Program of China(973 Program,No.2010CB428904,No.2011CB403606)Natural Science Foundation of China(No.41128006,No.40830854)
文摘Inter-annual variability of the Kuroshio water intrusion on the shelf of East China Sea (ECS) was simulated with a nested global and Northwest Pacific ocean circulation model.The model analysis reveals the influence of the variability of Kuroshio transport east of Taiwan on the intrusion to the northeast of Taiwan:high correlation (r=0.92) with the on-shore volume flux in the lower layer (50 200 m) ;low correlation (r=0.50) with the on-shore flux in the upper layer (0 50 m) .Spatial distribution of correlations between volume fluxes and sea surface height suggests that inter-annual variability of the Kuroshio flux east of Taiwan and its subsurface water intruding to the shelf lag behind the sea surface height anomalies in the central Pacific at 162 E by about 14 months,and could be related to wind-forced variation in the interior North Pacific that propagates westward as Rossby waves.The intrusion of Kuroshio surface water is also influenced by local winds.The intruding Kuroshio subsurface water causes variations of temperature and salinity of bottom waters on the southern ECS shelf.The influence of the intruding Kuroshio subsurface water extends widely from the shelf slope northeast of Taiwan northward to the central ECS near the 60 m isobath,and northeastward to the region near the 90 m isobath.
文摘The aim of the paper is to analyze a possible teleconnection of Quasi-Biennial Oscillation （QBO）, Southern Oscillation （SO）, North Atlantic Oscillation （NAO）, and Arctic Oscillation （AO） phenomena with longterm streamflow fluctuation of the Bela River （1895-2004） and Cierny Hron River （1931-2004） （central Slovakia）. Homogeneity, long-term trends, as well as inter-annual dry and wet cycles were analyzed for the entire 1895-2004 time series of the Bela River and for the 1931-2004 time series of the Cierny Hron River. Inter-annual fluctuation of the wet and dry periods was identified using spectral analysis. The most significant period is that of 3.6 years. Other significant periods are those of 2.35 years, 13.5 years, and 21 years. Since these periods were found in other rivers of the world, as well as in SO, NAO, and AO phenomena, they can be considered as relating to the general regularity of the Earth.
基金the National Key Research and Development Program of China(2016YFA0600203)the National Natural Science Foundation of China(41977191 and 41405138)the Major Programs of High-Resolution Earth Observation System(32-Y2-0A17-9001-15/17)。
文摘The study on how the variations in CO2 sources and sinks can affect the CO2 concentration over East Asia would be useful to provide information for policymaker concerning carbon emission reduction.In this study,a nested-grid version of global chemical transport model(GEOS-Chem)is employed to assess the impacts of variations in meteorological parameters,terrestrial fluxes,fossil fuel emissions,and biomass burning on inter-annual variations of CO2 concentrations over East Asia in 2004—2012.Simulated CO2 concentrations are compared with observations at 14 surface stations from the World Data Centre for Greenhouse Gases(WDCGG)and satellite-derived C 02 column density(XCO,)from the Gases Observing SATellite(GOSAT).The comparison shows that the simulated CO2 column density is generally higher than that of GOSAT by 1.33×10^6(annual mean point by point biases averaged over East Asia).The model reasonably captures the temporal variations of CO2 concentrations observed at the ground-based stations,but it is likely to underestimate the peaks-to-troughs amplitude of the seasonal cycle by 50%or more.The simulated surface CO2 concentration in East Asia exhibits the largest inter-annual variation in December-January—February(DJF).The regional mean absolute deviation(MAD)values over East Asia are within(4.4—5.0)×10^-6 for all seasons.Model sensitivity simulations indicate that the inter-annual variations of surface CO2 concentrations are mainly driven by variations of meteorological parameters,and partly modulated by the inter-annual variations of terrestrial fluxes and fossil fuel emissions in local regions.The variations of the terrestrial fluxes and fossil fuel emissions may account for〜28%of the inter-annual variation of surface CO2 concentration in southern China.The inter-annual variations of the peaks-to-troughs amplitude are dependent on variations of meteorological parameters,terrestrial fluxes and fossil fuel emissions in local regions.The influence of biomass burning emissions is relatively weak.
基金The Basic Research Operating Funds of The First Institute of Oceanography,State Oceanic Administration of China under contact Nos 2014T02 and 2014G02the Chinese Polar Environment Comprehensive Investigation and Assessment Programmes,State Oceanic Administration of China under contact Nos CHINARE2016-03-01 and CHINARE2016-04-03the Public Science and Technology Research Fund Project of Ocean under contact No.201205007
文摘On the basis of the CTD data obtained within the Bering Sea shelf by the Second to Sixth Chinese National Arctic Research Expedition in the summers of 2003, 2008, 2010, 2012 and 2014, the classification and interannual variation of water masses on the central Bering Sea shelf and the northern Bering Sea shelf are analyzed. The results indicate that there are both connection and difference between two regions in hydrological features. On the central Bering Sea shelf, there are mainly four types of water masses distribute orderly from the slope to the coast of Alaska： Bering Slope Current Water（BSCW）, MW（Mixed Water）, Bering Shelf Water（BSW） and Alaska Coastal Water（ACW）. In summer, BSW can be divided into Bering Shelf Surface Water（BSW_S） and Bering Shelf Cold Water（BSW_C）. On the northern Bering Sea shelf near the Bering Strait,it contains Anadyr Water（AW）, BSW and ACW from west to east. But the spatial-temporal features are also remarkable in each region. On the central shelf, the BSCW is saltiest and occupies the west of 177°W, which has the highest salinity in 2014. The BSW_C is the coldest water mass and warmest in 2014; the ACW is freshest and mainly occupies the east of 170°W, which has the highest temperature and salinity in 2012. On the northern Bering Sea shelf near the Bering Strait, the AW is saltiest with temperature decreasing sharply compared with BSCW on the central shelf. In the process of moving northward to the Bering Strait, the AW demonstrates a trend of eastward expansion. The ACW is freshest but saltier than the ACW on the central shelf,which is usually located above the BSW and is saltiest in 2014. The BSW distributes between the AW and the ACW and coldest in 2012, but the cold water of the BSW_C on the central shelf, whose temperature less than 0°C, does not exist on the northern shelf. Although there are so many changes, the respond to a climate change is synchronized in the both regions, which can be divided into the warm years（2003 and 2014） and cold years（2008, 2010 and 2012）. The year of 2014 may be a new beginning of warm period.
基金supported by the Natural Science Foundation of Shanghai (17ZR1435600)the Open Fund of Key Laboratory of Geospace Environment and Geodesy, Ministry of Education, Wuhan University (16-01-05)the National Key Research and Development Program of China (2016YFB0501405)
文摘Using more than 14 years of GRACE(Gravity Recovery and Climate Experiment) satellite gravimetry observations, we estimate the ice loss rate for the Patagonia Ice Field(PIF) of South America. After correcting the effects of glacier isostatic adjustment(GIA) and hydrological variations, the ice loss rate is -23.5 ± 8.1 Giga ton per year(Gt/yr) during the period April 2002 through December 2016, equivalent to an average ice thickness change of-1.3 m/yr if evenly distributed over PIF. The PIF ice mass change series also show obvious inter-annual variations during the entire period. For the time spans April 2002 to December 2007, January 2008 to December 2012 and January 2013 to December 2016, the ice loss rates are -26.4,-9.0 and -25.0 Gt/yr, respectively, indicating that the ice melting experienced significant slowing down and accelerating again in the past decade. Comparison with time series from temperature and precipitation data over PIF suggests that the inter-annual ice losses might not be directly correlated with the temperature changes and precipitation anomalies, and thus their interrelation is intricate. However, the dramatic ice loss acceleration in 2016(with more than 100 Gt within the first half of the year) appears closely related with the evident temperature increase and severe precipitation shortage over 2016, which are likely correlated with the strong E1 Nino event around 2016. Moreover, we compare the GRACE spherical harmonic(SH) and mass concentration(Mascon) solutions in estimating the PIF ice loss rate, and find that the Mascon result has larger uncertainty in leakage error correction,while the SH solutions can better correct leakage errors based on a constrained forward modeling iterative method. Thus the GRACE SH solutions with constrained forward modeling recovery are recommended to evaluating the ice mass change of PIF or other glacier regions with relatively smaller spatial scales.
文摘Plants play an essential role in matter and energy transformations and are key messengers in the carbon and energy cycle. Net primary productivity (NPP) reflects the capability of plants to transform solar energy into photosynthesis. It is very sensible for factors affecting on vegetation variability such as climate, soils, plant characteristics and human activities. So, it can be used as an indicator of actual and potential trend of vegetation. In this study we used the actual NPP which was derived from MODIS to assess the response of NPP to climate variables in Gadarif State, from 2000 to 2010. The correlations between NPP and climate variables (temperature and precipitation) are calculated using Pearson’s Correlation Coefficient and ordinary least squares regression. The main results show the following 1) the correlation Coefficient between NPP and mean annual temperature is Somewhat negative for Feshaga, Rahd, Gadarif and Galabat areas and weakly negative in Faw area;2) the correlation Coefficient between NPP and annual total precipitation is weakly negative in Faw, Rahd and Galabat areas and somewhat negative in Galabat and Rahd areas. This study demonstrated that the correlation analysis between NPP and climate variables (precipitation and temperature) gives reliably result of NPP responses to climate variables that is clearly in a very large scale of study area.
基金supported by the National Natural Science Foundation of China(No.51905123)Major Scientific and Technological Innovation Program of Shandong Province,China(Nos.2020CXGC010303,2022ZLGX04)Key R&D Programme of Shandong Province,China(No.2022JMRH0308).
文摘An internal state variable(ISV)model was established according to the experimental results of hot plane strain compression(PSC)to predict the microstructure evolution during hot spinning of ZK61 alloy.The effects of the internal variables were considered in this ISV model,and the parameters were optimized by genetic algorithm.After validation,the ISV model was used to simulate the evolution of grain size(GS)and dynamic recrystallization(DRX)fraction during hot spinning via Abaqus and its subroutine Vumat.By comparing the simulated results with the experimental results,the application of the ISV model was proven to be reliable.Meanwhile,the strength of the thin-walled spun ZK61 tube increased from 303 to 334 MPa due to grain refinement by DRX and texture strengthening.Besides,some ultrafine grains(0.5μm)that played an important role in mechanical properties were formed due to the proliferation,movement,and entanglement of dislocations during the spinning process.
文摘In order to solve the problem of the variable coefficient ordinary differen-tial equation on the bounded domain,the Lagrange interpolation method is used to approximate the exact solution of the equation,and the error between the numerical solution and the exact solution is obtained,and then compared with the error formed by the difference method,it is concluded that the Lagrange interpolation method is more effective in solving the variable coefficient ordinary differential equation.
基金supported by the National Natural Science Foundation China(Grant No.42176222).
文摘As a crucial component of the Earth’s climate system,Antarctic sea ice has demonstrated significant variability over the satellite era.Here,we identify a remarkable decadal transition in the total Antarctic Sea Ice Extent(SIE).The stage from 1979 to 2006 is characterized by high-frequency(i.e.,seasonal to interannual)temporal variability in SIE and zonal asymmetry in Sea Ice Concentration(SIC),which is primarily under the control of the Amundsen Sea Low(ASL).After 2007,however,sea ice changes exhibit a more spatially homogeneous pattern in SIC and a more temporally long-lasting mode in SIE.Further analysis reveals that sea ice-ocean interaction plays a major role in the low-frequency(i.e.,multiannual)variability of Antarctic sea ice from 2007−22.The related physical process is inferred to manifest as a strong coupling between the surface and the subsurface ocean layers,involving enhanced vertical convection and the downward delivery of the surface anomalies related to ice melting and freezing processes,thus maintaining the SIE anomalies for a longer time.Furthermore,this process mainly occurs in the Amundsen-Bellingshausen Sea(ABS)sector,and the weakened subsurface ocean stratification is the key factor triggering the coupling process in this region.We find that the Circumpolar Deep Water(CDW)over the ABS sector continued to shoal before 2007 and remained stable thereafter.It is speculated that the shoaling of the CDW may be a possible driver leading to the weakening of the subsurface stratification.
基金jointly supported by projects of the National Natural Science Foundation of China [grant numbers 42141017 and 41975112]。
文摘El Niño-Southern Oscillation(ENSO)is a major driver of climate change in middle and low latitudes and thus strongly influences the terrestrial carbon cycle through land-air interaction.Both the ENSO modulation and carbon flux variability are projected to increase in the future,but their connection still needs further investigation.To investigate the impact of future ENSO modulation on carbon flux variability,this study used 10 CMIP6 earth system models to analyze ENSO modulation and carbon flux variability in middle and low latitudes,and their relationship,under different scenarios simulated by CMIP6 models.The results show a high consistency in the simulations,with both ENSO modulation and carbon flux variability showing an increasing trend in the future.The higher the emissions scenario,especially SSP5-8.5 compared to SSP2-4.5,the greater the increase in variability.Carbon flux variability in the middle and low latitudes under SSP2-4.5 increases by 30.9%compared to historical levels during 1951-2000,while under SSP5-8.5 it increases by 58.2%.Further analysis suggests that ENSO influences mid-and low-latitude carbon flux variability primarily through temperature.This occurrence may potentially be attributed to the increased responsiveness of gross primary productivity towards regional temperature fluctuations,combined with the intensified influence of ENSO on land surface temperatures.
基金supported by the National Key R&D Program of China(No.2021YFB0301200)National Natural Science Foundation of China(No.62025208).
文摘Large-scale Language Models(LLMs)have achieved significant breakthroughs in Natural Language Processing(NLP),driven by the pre-training and fine-tuning paradigm.While this approach allows models to specialize in specific tasks with reduced training costs,the substantial memory requirements during fine-tuning present a barrier to broader deployment.Parameter-Efficient Fine-Tuning(PEFT)techniques,such as Low-Rank Adaptation(LoRA),and parameter quantization methods have emerged as solutions to address these challenges by optimizing memory usage and computational efficiency.Among these,QLoRA,which combines PEFT and quantization,has demonstrated notable success in reducing memory footprints during fine-tuning,prompting the development of various QLoRA variants.Despite these advancements,the quantitative impact of key variables on the fine-tuning performance of quantized LLMs remains underexplored.This study presents a comprehensive analysis of these key variables,focusing on their influence across different layer types and depths within LLM architectures.Our investigation uncovers several critical findings:(1)Larger layers,such as MLP layers,can maintain performance despite reductions in adapter rank,while smaller layers,like self-attention layers,aremore sensitive to such changes;(2)The effectiveness of balancing factors depends more on specific values rather than layer type or depth;(3)In quantization-aware fine-tuning,larger layers can effectively utilize smaller adapters,whereas smaller layers struggle to do so.These insights suggest that layer type is a more significant determinant of fine-tuning success than layer depth when optimizing quantized LLMs.Moreover,for the same discount of trainable parameters,reducing the trainable parameters in a larger layer is more effective in preserving fine-tuning accuracy than in a smaller one.This study provides valuable guidance for more efficient fine-tuning strategies and opens avenues for further research into optimizing LLM fine-tuning in resource-constrained environments.
基金supported by the National Natural Science Foundation of China [grant number 42088101]。
文摘During the boreal summer,intraseasonal oscillations exhibit significant interannual variations in intensity over two key regions:the central-western equatorial Pacific(5°S-5°N,150°E-150°W)and the subtropical Northwestern Pacific(10°-20°N,130°E-175°W).The former is well-documented and considered to be influenced by the ENSO,while the latter has received comparatively less attention and is likely influenced by the Pacific Meridional Mode(PMM),as suggested by partial correlation analysis results.To elucidate the physical processes responsible for the enhanced(weakened)intraseasonal convection over the subtropical northwestern Pacific during warm(cold)PMM years,the authors employed a moisture budget analysis.The findings reveal that during warm PMM years,there is an increase in summer-mean moisture over the subtropical northwestern Pacific.This increase interacts with intensified vertical motion perturbations in the region,leading to greater vertical moisture advection in the lower troposphere and consequently resulting in convective instability.Such a process is pivotal in amplifying intraseasonal convection anomalies.The observational findings were further verified by model experiments forced by PMM-like sea surface temperature patterns.
