Background/Context: Despite the high economic returns to college completion (Avery & Turner, 2012; Carnevale, Jayasundera, & Gulish, 2016), just over half of students who enroll at college attain a bachelor's degree (Bound, Lovenheim, & Turner, 2010; Denning, Eide, & Warnick, 2019; Shapiro et al., 2016). Colleges and non-profits have invested in various strategies to improve college completion, ranging from high-resource intensive advising (Scrivener et al., 2015) to light-touch messaging campaigns (Castleman & Page, 2016). More recently, organizations have implemented hybrid virtual advising models that provide continuous, light-touch supports as well as personalized advising to students after matriculation (Gurantz et al., 2020; Oreopoulos & Petronijevic, 2018; Page & Gehlbach, 2017; Sullivan et al., 2019). While these virtual advising strategies continue to grow in popularity, evaluation efforts have revealed wide variation in their effectiveness across contexts and implementations. One potential explanation for these mixed results is variability in the degree and type of student-advisor engagement that occurs, both within and across interventions. To the extent that students and advisors may navigate advising interactions in any variety of ways, students enrolled in virtual advising interventions may actually experience distinct treatments from one another. Moreover, the relationship between variation in these advising interactions and student outcomes is still ambiguous (Fesler, 2020). Text-based virtual advising programs offer an important opportunity to better examine these implementation dynamics, as nearly all student-advisor interactions are captured in the messaging data automatically collected by large-scale text messaging platforms. The availability of these relevant data, coupled with increasingly easy-to-implement text analysis techniques, allows researchers to better observe and measure how treatments vary across advising pairs while connecting them to outcomes of interest. Purpose/Objective/Research Question: In this study, we will conduct natural language processing (NLP) analyses on student-advisor interaction data from the Nudges to the Finish Line (N2FL) text-advising intervention to explore variability in advising interactions and program implementation (Bettinger, Castleman, & Mabel, 2019). N2FL was randomly offered to students approaching degree completion at over twenty colleges and universities. Students in the intervention then received pre-scheduled informational messages regarding important deadlines and campus resources, as well as opportunities to engage in personalized conversations with advisors about their academic progress. Population/Participants: Our sample includes 6,602 treated students assigned across 77 advisors. On average, students in the sample had completed about 57 credits and overall 28 percent had "stopped out" (unenrolled from the college without graduating), though this varied substantially from 17 percent in State A's institutions to 38 percent at State B's institutions. The average age at entry into college was around 21; note there are substantive racial/ethnic differences across the two college system populations, given demographic differences between the two geographies. About a fifth of the State B population identifies as White relative to nine percent of the State A population where about two fifths of students identify as Black and 30 percent identify as Hispanic. Research Design: For the present study, we quantify variation in the advising each student experienced throughout the intervention by conducting descriptive analysis to document message intensity (i.e. number and length), advanced sentiment analysis to measure message tone (i.e. positivity/negativity of ideas and emotions), and structural topic modeling to describe message content (i.e. discussions of financial aid versus course registration). We then go on to examine how variation in these measures correlate with the intervention's outcomes of interest -- credit accumulation and college completion -- after controlling for a rich set of student baseline covariates. Findings/Results: In our preliminary analyses, we find sentiment variation in advisor-initiated and student-initiated messages. Approximately, 30% of student-initiated texts were negative in tone, compared with 16% of advisor-initiated messages, likely because students are more likely to reach out when struggling or confused, resulting in texts that pertain to negative ideas like failing or financial difficulties. Our topic modeling analysis reveals that relative to the scheduled messages, students and advisors are more likely to discuss advising scheduling logistics and less likely to discuss financial aid (13 percent of scheduled messages discuss financial aid relative to 5-7 percent of impromptu student-advisor conversations). This is an important comparison for institutions to keep in mind as they understand what topics students have more questions about to inform their development of future scheduled messages. Relationships between measured sentiment and topics and student outcomes are ongoing. Conclusions: Our findings offer two main contributions to the field. First, we demonstrate how text-as-data methods can be carefully and robustly applied to contextualize the findings of multi-site, text-based field experiments. Preliminary findings for the N2FL intervention revealed no overall impacts of the text advising campaign on student outcomes such as persistence and graduation, but found suggestive evidence for larger subgroup impacts among students at higher risk of withdrawal prior to completing their degree program. Variation in implementation across sites may be a key factor in explaining some of this treatment heterogeneity, and we will offer a proof-of-concept for how such variation could be systematically measured with NLP to better understand treatment heterogeneity in experimental contexts. Second, this analysis more specifically offers timely insight into the nature and content of one-on-one interactions between students and advisors. Since the COVID-19 pandemic began, an increasing number of higher education institutions have turned to remote advising practices and campus-wide text messaging campaigns to support their students. This study will provide important evidence for how student, advisor, and institutional characteristics contribute to the shape of advising practices, informing higher education institutions' adoption and implementation of these remote advising programs going forward.