Search for: All records

Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries. 

Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries. 

Abstract We present the Local GroupL-Band Survey, a Karl G. Jansky Very Large Array (VLA) survey producing the highest-quality 21 cm and 1–2 GHz radio continuum images to date, for the six VLA-accessible, star-forming, Local Group galaxies. Leveraging the VLA’s spectral multiplexing power, we simultaneously survey the 21 cm line at high 0.4 km s⁻¹velocity resolution, the 1–2 GHz polarized continuum, and four OH lines. For the massive spiral M31, the dwarf spiral M33, and the dwarf irregular galaxies NGC 6822, IC 10, IC 1613, and the Wolf–Lundmark–Melotte Galaxy, we use all four VLA configurations and the Green Bank Telescope to reach angular resolutions of <5″ (10–20 pc) for the 21 cm line with <10²⁰cm⁻²column density sensitivity, and even sharper views (<2″; 5–10 pc) of the continuum. Targeting these nearby galaxies (D ≲ 1 Mpc) reveals a sharp, resolved view of the atomic gas, including 21 cm absorption, and continuum emission from supernova remnants and Hiiregions. These data sets can be used to test theories of the abundance and formation of cold clouds, the driving and dissipation of interstellar turbulence, and the impact of feedback from massive stars and supernovae. Here, we describe the survey design and execution, scientific motivation, data processing, and quality assurance. We provide a first look at and publicly release the wide-field 21 cm Hidata products for M31, M33, and four dwarf irregular targets in the survey, which represent some of the highest-physical-resolution 21 cm observations of any external galaxies beyond the LMC and SMC. 

Abstract Feedback from star formation is a critical component of the evolution of galaxies and their interstellar medium. At parsec scales internal to molecular clouds, however, the observed signatures of that feedback on the physical properties of CO-emitting gas have often been weak or inconclusive. We present subparsec observations of H₂CO in the 30 Doradus region, which contains the massive star cluster R136 that is clearly exerting feedback on its neighboring gas. H₂CO provides a direct measure of gas kinetic temperature, and we find a trend of decreasing temperature with projected distance from R136 that may be indicative of gas heating by the stars. While it has been suggested that mechanical heating affects H₂CO-measured temperature, we do not observe any correlation betweenT_Kand line width. The lack of an enhancement in mechanical feedback close to R136 is consistent with the absence of a radial trend in gravitational boundedness seen the Atacama Large Millimeter/submillimeter Array CO observations. Estimates of cosmic-ray flux in the region are quite uncertain, but can plausibly explain the observed temperatures if R136 itself is the dominant local source of energetic protons. The observations presented here are also consistent with the H₂CO-emitting gas near R136 being dominated by direct radiation from R136 and photoelectric heating in the photodissociation regions. 

Graduation rates are a key measure of the long-term efficacy of academic interventions. However, challenges to using traditional estimates of graduation rates for underrepresented students include inherently small sample sizes and high data requirements. Here, we show that a Markov model increases confidence and reduces biases in estimated graduation rates for underrepresented minority and first-generation students. We use a Learning Assistant program to demonstrate the Markov model’s strength for assessing program efficacy. We find that Learning Assistants in gateway science courses are associated with a 9% increase in the six-year graduation rate. These gains are larger for underrepresented minority (21%) and first-generation students (18%). Our results indicate that Learning Assistants can improve overall graduation rates and address inequalities in graduation rates for underrepresented students. 

The Magellanic Stream (MS), a tail of diffuse gas formed from tidal and ram pressure interactions between the Small and Large Magellanic Clouds (SMC and LMC) and the Halo of the Milky Way, is primarily composed of neutral atomic hydrogen (HI). The deficiency of dust and the diffuse nature of the present gas make molecular formation rare and difficult, but if present, could lead to regions potentially suitable for star formation, thereby allowing us to probe conditions of star formation similar to those at high redshifts. We search for HCO+ ,HCN,HNC,andC2H using the highest sensitivity observations of molecular absorption data from the Atacama Large Millimeter Array (ALMA) to trace these regions, comparing with HI archival data from the Galactic Arecibo L-Band Feed Array (GALFA) HI Survey and the Galactic All Sky Survey (GASS) to compare these environments in the MS to the HI column density threshold for molecular formation in the Milky Way. We also compare the line of sight locations with confirmed locations of stars, molecular hydrogen, and OI detections, though at higher sensitivities than the observations presented here.