Showing 1–7 of 7 results for author: Perez-Becker, D

Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

Authors: Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sébastien Bubeck, Qin Cai, Martin Cai, Caio César Teodoro Mendes, Weizhu Chen, Vishrav Chaudhary, Dong Chen, Dongdong Chen, Yen-Chun Chen, Yi-Ling Chen, Parul Chopra , et al. (90 additional authors not shown)

Abstract: We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. The innovation lies entirely in our dataset… ▽ More We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. The innovation lies entirely in our dataset for training, a scaled-up version of the one used for phi-2, composed of heavily filtered publicly available web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide some initial parameter-scaling results with a 7B and 14B models trained for 4.8T tokens, called phi-3-small and phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75% and 78% on MMLU, and 8.7 and 8.9 on MT-bench). Moreover, we also introduce phi-3-vision, a 4.2 billion parameter model based on phi-3-mini with strong reasoning capabilities for image and text prompts. △ Less

Submitted 23 May, 2024; v1 submitted 22 April, 2024; originally announced April 2024.

Comments: 19 pages

Atmospheric Heat Redistribution on Hot Jupiters

Authors: Daniel Perez-Becker, Adam P. Showman

Abstract: Infrared lightcurves of transiting hot Jupiters present a trend in which the atmospheres of the hottest planets are less efficient at redistributing the stellar energy absorbed on their daysides---and thus have a larger day-night temperature contrast---than colder planets. No predictive atmospheric model has been published that identifies which dynamical mechanisms determine the atmospheric heat r… ▽ More Infrared lightcurves of transiting hot Jupiters present a trend in which the atmospheres of the hottest planets are less efficient at redistributing the stellar energy absorbed on their daysides---and thus have a larger day-night temperature contrast---than colder planets. No predictive atmospheric model has been published that identifies which dynamical mechanisms determine the atmospheric heat redistribution efficiency on tidally locked exoplanets. Here we present a two-layer shallow water model of the atmospheric dynamics on synchronously rotating planets that explains the observed trend. Our model shows that planets with weak friction and weak irradiation exhibit a banded zonal flow with minimal day-night temperature differences, while models with strong irradiation and/or strong friction exhibit a day-night flow pattern with order-unity fractional day-night temperature differences. To interpret the model, we develop a scaling theory that shows that the timescale for gravity waves to propagate horizontally over planetary scales, t_wave, plays a dominant role in controlling the transition from small to large temperature contrasts. This implies that heat redistribution is governed by a wave-like process, similar to the one responsible for the weak temperature gradients in the Earth's tropics. When atmospheric drag can be neglected, the transition from small to large day-night temperature contrasts occurs when t_wave ~ sqrt(t_rad/Omega), where t_rad is the radiative relaxation time and Omega is the planetary rotation frequency. Alternatively, this transition criterion can be expressed as t_rad ~ t_vert, where t_vert is the timescale for a fluid parcel to move vertically over the difference in day-night thickness. These results subsume the commonly used timescale comparison for estimating heat redistribution efficiency between t_rad and the global horizontal advection timescale, t_adv. △ Less

Submitted 25 September, 2013; v1 submitted 19 June, 2013; originally announced June 2013.

Comments: Accepted to ApJ with minor edits compared to version 1; 17 pages, 11 figures

Catastrophic Evaporation of Rocky Planets

Authors: Daniel Perez-Becker, Eugene Chiang

Abstract: Short-period exoplanets can have dayside surface temperatures surpassing 2000 K, hot enough to vaporize rock and drive a thermal wind. Small enough planets evaporate completely. We construct a radiative-hydrodynamic model of atmospheric escape from strongly irradiated, low-mass rocky planets, accounting for dust-gas energy exchange in the wind. Rocky planets with masses < 0.1 M_Earth (less than tw… ▽ More Short-period exoplanets can have dayside surface temperatures surpassing 2000 K, hot enough to vaporize rock and drive a thermal wind. Small enough planets evaporate completely. We construct a radiative-hydrodynamic model of atmospheric escape from strongly irradiated, low-mass rocky planets, accounting for dust-gas energy exchange in the wind. Rocky planets with masses < 0.1 M_Earth (less than twice the mass of Mercury) and surface temperatures > 2000 K are found to disintegrate entirely in < 10 Gyr. When our model is applied to Kepler planet candidate KIC 12557548b --- which is believed to be a rocky body evaporating at a rate of dM/dt > 0.1 M_Earth/Gyr --- our model yields a present-day planet mass of < 0.02 M_Earth or less than about twice the mass of the Moon. Mass loss rates depend so strongly on planet mass that bodies can reside on close-in orbits for Gyrs with initial masses comparable to or less than that of Mercury, before entering a final short-lived phase of catastrophic mass loss (which KIC 12557548b has entered). Because this catastrophic stage lasts only up to a few percent of the planet's life, we estimate that for every object like KIC 12557548b, there should be 10--100 close-in quiescent progenitors with sub-day periods whose hard-surface transits may be detectable by Kepler --- if the progenitors are as large as their maximal, Mercury-like sizes (alternatively, the progenitors could be smaller and more numerous). According to our calculations, KIC 12557548b may have lost ~70% of its formation mass; today we may be observing its naked iron core. △ Less

Submitted 21 May, 2013; v1 submitted 8 February, 2013; originally announced February 2013.

Comments: Accepted to MNRAS with minor edits compared to version 2

Detection and Imaging of the Crab Nebula with the Nuclear Compton Telescope

Authors: M. S. Bandstra, E. C. Bellm, S. E. Boggs, D. Perez-Becker, A. Zoglauer, H. -K. Chang, J. -L. Chiu, J. -S. Liang, Y. -H. Chang, Z. -K. Liu, W. -C. Hung, M. -H. A. Huang, S. J. Chiang, R. -S. Run, C. -H. Lin, M. Amman, P. N. Luke, P. Jean, P. von Ballmoos, C. B. Wunderer

Abstract: The Nuclear Compton Telescope (NCT) is a balloon-borne Compton telescope designed for the study of astrophysical sources in the soft gamma-ray regime (200 keV--20 MeV). NCT's ten high-purity germanium crossed-strip detectors measure the deposited energies and three-dimensional positions of gamma-ray interactions in the sensitive volume, and this information is used to restrict the initial photon t… ▽ More The Nuclear Compton Telescope (NCT) is a balloon-borne Compton telescope designed for the study of astrophysical sources in the soft gamma-ray regime (200 keV--20 MeV). NCT's ten high-purity germanium crossed-strip detectors measure the deposited energies and three-dimensional positions of gamma-ray interactions in the sensitive volume, and this information is used to restrict the initial photon to a circle on the sky using the Compton scatter technique. Thus NCT is able to perform spectroscopy, imaging, and polarization analysis on soft gamma-ray sources. NCT is one of the next generation of Compton telescopes --- so-called compact Compton telescopes (CCTs) --- which can achieve effective areas comparable to COMPTEL's with an instrument that is a fraction of the size. The Crab Nebula was the primary target for the second flight of the NCT instrument, which occurred on 17--18 May 2009 in Fort Sumner, New Mexico. Analysis of 29.3 ks of data from the flight reveals an image of the Crab at a significance of 4-sigma. This is the first reported detection of an astrophysical source by a CCT. △ Less

Submitted 1 June, 2011; originally announced June 2011.

Comments: 10 pages, 9 figures, accepted for publication in the Astrophysical Journal

Surface Layer Accretion in Conventional and Transitional Disks Driven by Far-Ultraviolet Ionization

Authors: Daniel Perez-Becker, Eugene Chiang

Abstract: Whether protoplanetary disks accrete at observationally significant rates by the magnetorotational instability (MRI) depends on how well ionized they are. Disk surface layers ionized by stellar X-rays are susceptible to charge neutralization by small condensates, ranging from ~0.01-micron-sized grains to angstrom-sized polycyclic aromatic hydrocarbons (PAHs). Ion densities in X-ray-irradiated surf… ▽ More Whether protoplanetary disks accrete at observationally significant rates by the magnetorotational instability (MRI) depends on how well ionized they are. Disk surface layers ionized by stellar X-rays are susceptible to charge neutralization by small condensates, ranging from ~0.01-micron-sized grains to angstrom-sized polycyclic aromatic hydrocarbons (PAHs). Ion densities in X-ray-irradiated surfaces are so low that ambipolar diffusion weakens the MRI. Here we show that ionization by stellar far-ultraviolet (FUV) radiation enables full-blown MRI turbulence in disk surface layers. Far-UV ionization of atomic carbon and sulfur produces a plasma so dense that it is immune to ion recombination on grains and PAHs. The FUV-ionized layer, of thickness 0.01--0.1 g/cm^2, behaves in the ideal magnetohydrodynamic limit and can accrete at observationally significant rates at radii > 1--10 AU. Surface layer accretion driven by FUV ionization can reproduce the trend of increasing accretion rate with increasing hole size seen in transitional disks. At radii < 1--10 AU, FUV-ionized surface layers cannot sustain the accretion rates generated at larger distance, and unless turbulent mixing of plasma can thicken the MRI-active layer, an additional means of transport is needed. In the case of transitional disks, it could be provided by planets. △ Less

Submitted 9 May, 2011; v1 submitted 12 April, 2011; originally announced April 2011.

Comments: Final proofed version. Corrects X-ray-driven accretion rates in the high PAH case for Figures 8 and 9

Surface Layer Accretion in Transitional and Conventional Disks: From Polycyclic Aromatic Hydrocarbons to Planets

Authors: Daniel Perez-Becker, Eugene Chiang

Abstract: Transitional T Tauri disks have optically thin holes with radii > 10 AU, yet accrete up to the median T Tauri rate. Multiple planets inside the hole can torque the gas to high radial speeds over large distances, reducing the local surface density while maintaining accretion. Thus multi-planet systems, together with reductions in disk opacity due to grain growth, can explain how holes can be simult… ▽ More Transitional T Tauri disks have optically thin holes with radii > 10 AU, yet accrete up to the median T Tauri rate. Multiple planets inside the hole can torque the gas to high radial speeds over large distances, reducing the local surface density while maintaining accretion. Thus multi-planet systems, together with reductions in disk opacity due to grain growth, can explain how holes can be simultaneously transparent and accreting. There remains the problem of how outer disk gas diffuses into the hole. Here it has been proposed that the magnetorotational instability (MRI) erodes disk surface layers ionized by stellar X-rays. In contrast to previous work, we find that the extent to which surface layers are MRI-active is limited not by ohmic dissipation but by ambipolar diffusion, the latter measured by Am: the number of times a neutral hydrogen molecule collides with ions in a dynamical time. Simulations by Hawley & Stone showed that Am ~ 100 is necessary for ions to drive MRI turbulence in neutral gas. We calculate that in X-ray-irradiated surface layers, Am typically varies from ~0.001 to 1, depending on the abundance of charge-adsorbing polycyclic aromatic hydrocarbons, whose properties we infer from Spitzer observations. We conclude that ionization of H_2 by X-rays and cosmic rays can sustain, at most, only weak MRI turbulence in surface layers 1 to 10 g/cm^2 thick, and that accretion rates in such layers are too small compared to observed accretion rates for the majority of disks. △ Less

Submitted 7 November, 2010; v1 submitted 24 September, 2010; originally announced September 2010.

Comments: Accepted to ApJ. Final version. Note added in proof where accretion rate estimates are given

The 2.3 GHz continuum survey of the GEM project

Authors: C. Tello, T. Villela, S. Torres, M. Bersanelli, G. F. Smoot, I. S. Ferreira, A. Cingoz, J. Lamb, D. Barbosa, D. Perez-Becker, S. Ricciardi, J. A. Currivan, P. Platania, D. Maino

Abstract: We present a partial-sky survey of the radio continuum at 2.3 GHz within the scope of the Galactic Emission Map** (GEM) project, an observational program conceived and developed to reveal the large-scale properties of Galactic synchrotron radiation through a set of self-consistent surveys of the radio continuum between 408 MHz and 10 GHz. The GEM experiment uses a portable and double-shielded 5.… ▽ More We present a partial-sky survey of the radio continuum at 2.3 GHz within the scope of the Galactic Emission Map** (GEM) project, an observational program conceived and developed to reveal the large-scale properties of Galactic synchrotron radiation through a set of self-consistent surveys of the radio continuum between 408 MHz and 10 GHz. The GEM experiment uses a portable and double-shielded 5.5-m radiotelescope in altazimuthal configuration to map 60-degree-wide declination bands from different observational sites by circularly scanning the sky at zenithal angles of 30 deg from a constantly rotating platform. The observations were accomplished with a total power receiver, whose front-end High Electron Mobility Transistor (HEMT) amplifier was matched directly to a cylindrical horn at the prime focus of the parabolic reflector. The Moon was used to calibrate the antenna temperature scale and the preparation of the map required direct subtraction and destri** algorithms to remove ground contamination as the most significant source of systematic error. We used 484 hours of total intensity observations from two locations in Colombia and Brazil to yield 66% sky coverage from DEC = -51.73 deg to DEC = +34.78 deg. The zero-level uncertainty of the combined survey is 103 mK with a temperature scale error of 5% after direct correlation with the Rhodes/HartRAO survey at 2326 MHz on a T-T plot. The sky brightness distribution into regions of low and high emission in the GEM survey is consistent with the appearance of a transition region as seen in the Haslam 408 MHz and WMAP K-band surveys. Preliminary results also show that the temperature spectral index between 408 MHz and the 2.3 GHz band of the GEM survey has a weak spatial correlation with these regions; but it steepens significantly from high to low emission regions with respect to the WMAP K-band survey. △ Less

Submitted 7 July, 2013; v1 submitted 19 December, 2007; originally announced December 2007.

Comments: 20 pages, 21 figures, 6 tables. Extensively revised and enlarged version accepted for publication in Astronomy & Astrophysics. Smaller figures