Traian Rebedea
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Traian Rebedea has a total of 80 co-authors. Here are the top co-authors by number of shared publications:
- Stefan Trausan-Matu
Department of Computer Science and Engineering, Politehnica Bucharest, National University for Science and Technology, 060042 Bucharest, Romania
10 shared publications - Mihai Dascalu
Computer Science & Engineering Department, National University of Science and Technology POLITEHNICA Bucharest, 313 Splaiul Independentei, 060042 Bucharest, Romania
9 shared publications - Christopher Parisien
School of Computing and AI Arizona State University
6 shared publications - Makesh Narsimhan Sreedhar
NVIDIA Santa Clara CA
6 shared publications - Costin-Gabriel Chiru
Department of Computer Science, National University of Science and Technology Politehnica Bucharest, Bucharest, Romania
4 shared publications - Mihai Masala
Computer Science and Engineering Department, National University of Science and Technology POLITEHNICA Bucharest, 313 Splaiul Independentei, 060042 Bucharest, Romania
4 shared publications
Latest Publications
- Open Access
- 8 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 11052
Solar Forecasting: Small-Window Performance and Cross-Site Validation
Prosumers—i.e., households and small businesses that both produce and consume electricity—need short-term forecasts of their photovoltaic (PV) generation to plan storage, self-consumption and grid export, but rarely have the years of history that data-hungry deep models assume. We integrated seven forecasting models into our open-source toolkit, including gradient-boosted trees and recurrent, convolutional, attention-based, and basis-expansion networks; all models were evaluated on a newly released dataset from two grid-connected PV systems with 50 kW inverter ratings in southwestern Romania. Under a rolling-origin protocol in which every input, including day-ahead weather forecasts, is available when the forecast is issued, LightGBM and XGBoost reach 9.9–10.0% normalized RMSE one hour ahead, a skill of 7–9% over smart persistence, which they surpass in 11 of the 12 test months, whereas every deep model falls short of that reference, the best (a Transformer) reaching 14.2%. The boosted models perform better at all tested history lengths, from one month of data where the best reaches 12.6% against 16.9% for the best deep model, to 12 months where the best deep model still has 1.28 times the error of the best boosted one. The model ranking is identical at a second installation in the same municipality (Spearman ). In fixed-origin forecasts for the following days at the first site, day-ahead weather forecasts reduce the first-day error from 23.0% to 19.2%, and the benefit declines steadily with lead time, with no cutoff within the 7 days covered by the forecast archive.- 8 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 11052
Open AccessSolar Forecasting: Small-Window Performance and Cross-Site Validation
Prosumers—i.e., households and small businesses that both produce and consume electricity—need short-term forecasts of their photovoltaic (PV) generation to plan storage, self-consumption and grid export, but rarely have the years of history that data-hungry deep models assume. We integrated seven forecasting models into our open-source toolkit, including gradient-boosted trees and recurrent, convolutional, attention-based, and basis-expansion networks; all models were evaluated on a newly released dataset from two grid-connected PV systems with 50 kW inverter ratings in southwestern Romania. Under a rolling-origin protocol in which every input, including day-ahead weather forecasts, is available when the forecast is issued, LightGBM and XGBoost reach 9.9–10.0% normalized RMSE one hour ahead, a skill of 7–9% over smart persistence, which they surpass in 11 of the 12 test months, whereas every deep model falls short of that reference, the best (a Transformer) reaching 14.2%. The boosted models perform better at all tested history lengths, from one month of data where the best reaches 12.6% against 16.9% for the best deep model, to 12 months where the best deep model still has 1.28 times the error of the best boosted one. The model ranking is identical at a second installation in the same municipality (Spearman ). In fixed-origin forecasts for the following days at the first site, day-ahead weather forecasts reduce the first-day error from 23.0% to 19.2%, and the benefit declines steadily with lead time, with no cutoff within the 7 days covered by the forecast archive. - Open Access
Evaluating Safety Embedding Prefiltering for Analyzing Millions of LLM Agent Social Network Messages for Security and Safety Harms
Agentic social networks expose autonomous agents to large volumes of potentially malicious content, including prompt injection, social engineering, and unsafe execution requests. Large Language Model (LLM) judges can detect such content, but applying them at platform scale is expensive. We study the extent to which lightweight embedding-based prefilters can reduce this cost while retaining most of the judge’s unsafe predictions. We annotate 10,000 Moltbook posts and comments with a frontier LLM judge, assigning a binary safety verdict, severity level, malicious intent taxonomy labels, and Open Worldwide Application Security Project (OWASP) AI risk codes. Of the LLM-annotated samples, 9.2% were unsafe at severity 3 or above. Two human annotators with high inter-rater agreement (Cohen’s ) exhibited moderate agreement between their adjudicated labels and the LLM ratings (). We compare centroid-based cosine prefilters over 3 off-the-shelf encoders against a supervised trained classifier. Reducing the input window size raises average precision by up to 6.3 percentage points. At an operating point calibrated to 0.80 recall, the projected cost of scanning 787,226 messages falls from $7085 by 49.9% (MiniLM-L12-v2), 55.6% (BGE-M3), and 65.6% (trained classifier), while MiniLM runs roughly 32 times faster than the classifier. Prefiltering can therefore make frontier-judge screening affordable at scale, although jailbreak content remains the hardest category to retrieve.Open AccessEvaluating Safety Embedding Prefiltering for Analyzing Millions of LLM Agent Social Network Messages for Security and Safety Harms
Agentic social networks expose autonomous agents to large volumes of potentially malicious content, including prompt injection, social engineering, and unsafe execution requests. Large Language Model (LLM) judges can detect such content, but applying them at platform scale is expensive. We study the extent to which lightweight embedding-based prefilters can reduce this cost while retaining most of the judge’s unsafe predictions. We annotate 10,000 Moltbook posts and comments with a frontier LLM judge, assigning a binary safety verdict, severity level, malicious intent taxonomy labels, and Open Worldwide Application Security Project (OWASP) AI risk codes. Of the LLM-annotated samples, 9.2% were unsafe at severity 3 or above. Two human annotators with high inter-rater agreement (Cohen’s ) exhibited moderate agreement between their adjudicated labels and the LLM ratings (). We compare centroid-based cosine prefilters over 3 off-the-shelf encoders against a supervised trained classifier. Reducing the input window size raises average precision by up to 6.3 percentage points. At an operating point calibrated to 0.80 recall, the projected cost of scanning 787,226 messages falls from $7085 by 49.9% (MiniLM-L12-v2), 55.6% (BGE-M3), and 65.6% (trained classifier), while MiniLM runs roughly 32 times faster than the classifier. Prefiltering can therefore make frontier-judge screening affordable at scale, although jailbreak content remains the hardest category to retrieve. - Open Access
- 15 September 2026
- Published by MDPI AG in Preprints.org
Solar Forecasting: Small-Window Performance and Cross-Site Validation
Accurate short-term forecasting of photovoltaic (PV) generation, long a concern for grid operators, is increasingly relevant to prosumers as well: the households and small businesses that both produce and consume electricity and must plan storage, self-consumption, and grid export. This paper studies solar forecasting under the constraint that users rarely possess years of historical generation data that data-hungry deep models assume. We release an open-source toolkit that integrates 7 forecasting models, spanning gradient-boosted trees, recurrent and convolutional sequence models, and attention- and basis-expansion architectures. We evaluate the toolkit’s performance on a newly released dataset of two grid-connected 50 kWp PV systems in southwestern Romania, using a strict rolling-origin walk-forward protocol. Across all experiments, XGBoost and LightGBM lead: on the primary site, they achieve 10.4% normalized RMSE against 12.9% for the best deep model, and they already beat a naive persistence baseline after a single month of training data. A dedicated small-window study shows that the boosting advantage is concentrated in the cold-start regime and narrows steadily as history accumulates, from a factor of 2.1 at one month to 1.08 at twelve, with no model saturating inside the available record. A cross-site replication on a second, independent installation reproduces the full model ranking almost exactly (Spearman ρ=0.964), strong evidence that the finding is not a single-dataset artifact. The resulting forecaster is deliberately lightweight and cost-efficient: a few hundred shallow trees over a seven-feature lag/rolling core that trains in under a second per month on a commodity CPU. We publicly release the toolkit and the anonymized two-site dataset to support reproducible prosumer-scale PV forecasting research.
- 15 September 2026
- Published by MDPI AG in Preprints.org
Open AccessSolar Forecasting: Small-Window Performance and Cross-Site Validation
Accurate short-term forecasting of photovoltaic (PV) generation, long a concern for grid operators, is increasingly relevant to prosumers as well: the households and small businesses that both produce and consume electricity and must plan storage, self-consumption, and grid export. This paper studies solar forecasting under the constraint that users rarely possess years of historical generation data that data-hungry deep models assume. We release an open-source toolkit that integrates 7 forecasting models, spanning gradient-boosted trees, recurrent and convolutional sequence models, and attention- and basis-expansion architectures. We evaluate the toolkit’s performance on a newly released dataset of two grid-connected 50 kWp PV systems in southwestern Romania, using a strict rolling-origin walk-forward protocol. Across all experiments, XGBoost and LightGBM lead: on the primary site, they achieve 10.4% normalized RMSE against 12.9% for the best deep model, and they already beat a naive persistence baseline after a single month of training data. A dedicated small-window study shows that the boosting advantage is concentrated in the cold-start regime and narrows steadily as history accumulates, from a factor of 2.1 at one month to 1.08 at twelve, with no model saturating inside the available record. A cross-site replication on a second, independent installation reproduces the full model ranking almost exactly (Spearman ρ=0.964), strong evidence that the finding is not a single-dataset artifact. The resulting forecaster is deliberately lightweight and cost-efficient: a few hundred shallow trees over a seven-feature lag/rolling core that trains in under a second per month on a commodity CPU. We publicly release the toolkit and the anonymized two-site dataset to support reproducible prosumer-scale PV forecasting research.
- Open Access
- 4 September 2026
- Published by MDPI AG in Preprints.org
Evaluating Safety Embedding Prefiltering for Analyzing Millions of LLM Agent Social Network Messages for Security and Safety Harms
Agentic social networks expose autonomous agents to large volumes of potentially malicious content, including prompt injection, social engineering, and unsafe execution requests. Large Language Model (LLM) judges can detect such content, but applying them at platform scale is expensive. We study the extent to which lightweight embedding-based prefilters can reduce this cost while retaining most of the judge’s unsafe predictions. We annotate 10,000 Moltbook posts and comments with a frontier LLM judge, assigning a binary safety verdict, severity level, malicious intent taxonomy labels, and Open Worldwide Application Security Project (OWASP) AI risk codes. Of the LLM-annotated samples, 9.1% were unsafe at severity 3 or above. Two human annotators with high inter-rater agreement (Cohen’s κ = 0.823) exhibited moderate agreement between their adjudicated labels and the LLM ratings (κ = 0.578). We compare centroid-based cosine prefilters over 3 off-the-shelf encoders against a supervised trained classifier. Reducing the input window size raises average precision by up to 6.3 percentage points. At an operating point calibrated to 0.80 recall, the projected cost of scanning 787,226 messages falls from $7,085 by 56.0% (MiniLM-L12-v2), 62.6% (BGE-M3), and 74.0% (trained classifier), while MiniLM runs roughly 32 times faster than the classifier. Prefiltering can therefore make frontier-judge screening affordable at scale, although jailbreak content remains the hardest category to retrieve.
- 4 September 2026
- Published by MDPI AG in Preprints.org
Open AccessEvaluating Safety Embedding Prefiltering for Analyzing Millions of LLM Agent Social Network Messages for Security and Safety Harms
Agentic social networks expose autonomous agents to large volumes of potentially malicious content, including prompt injection, social engineering, and unsafe execution requests. Large Language Model (LLM) judges can detect such content, but applying them at platform scale is expensive. We study the extent to which lightweight embedding-based prefilters can reduce this cost while retaining most of the judge’s unsafe predictions. We annotate 10,000 Moltbook posts and comments with a frontier LLM judge, assigning a binary safety verdict, severity level, malicious intent taxonomy labels, and Open Worldwide Application Security Project (OWASP) AI risk codes. Of the LLM-annotated samples, 9.1% were unsafe at severity 3 or above. Two human annotators with high inter-rater agreement (Cohen’s κ = 0.823) exhibited moderate agreement between their adjudicated labels and the LLM ratings (κ = 0.578). We compare centroid-based cosine prefilters over 3 off-the-shelf encoders against a supervised trained classifier. Reducing the input window size raises average precision by up to 6.3 percentage points. At an operating point calibrated to 0.80 recall, the projected cost of scanning 787,226 messages falls from $7,085 by 56.0% (MiniLM-L12-v2), 62.6% (BGE-M3), and 74.0% (trained classifier), while MiniLM runs roughly 32 times faster than the classifier. Prefiltering can therefore make frontier-judge screening affordable at scale, although jailbreak content remains the hardest category to retrieve.
- Open Access
Semi-Supervised Learning for Large Language Models Safety and Content Moderation
Safety for Large Language Models (LLMs) has been an ongoing research focus since their emergence and is even more relevant nowadays with the increasing capacity of those models. Currently, there are several guardrails in place for all public LLMs and multiple proposed datasets for training safety classifiers. However, training these safety classifiers relies on large quantities of labeled data, which can be problematic to acquire, prone to labeling errors, or often include synthetic data. To address these issues, we suggest a different approach: utilizing semi-supervised learning techniques, which leverage both labeled and unlabeled data, to improve the performance on the safety task. We analyze the improvements that these techniques can offer for both prompts given to Large Language Models and the responses to those requests. Moreover, since augmentation is the central part of semi-supervised algorithms, we demonstrate the importance of using task-specific augmentations, which significantly increase the performance when compared to general-purpose augmentation techniques.
Open AccessSemi-Supervised Learning for Large Language Models Safety and Content Moderation
Safety for Large Language Models (LLMs) has been an ongoing research focus since their emergence and is even more relevant nowadays with the increasing capacity of those models. Currently, there are several guardrails in place for all public LLMs and multiple proposed datasets for training safety classifiers. However, training these safety classifiers relies on large quantities of labeled data, which can be problematic to acquire, prone to labeling errors, or often include synthetic data. To address these issues, we suggest a different approach: utilizing semi-supervised learning techniques, which leverage both labeled and unlabeled data, to improve the performance on the safety task. We analyze the improvements that these techniques can offer for both prompts given to Large Language Models and the responses to those requests. Moreover, since augmentation is the central part of semi-supervised algorithms, we demonstrate the importance of using task-specific augmentations, which significantly increase the performance when compared to general-purpose augmentation techniques.
- Open Access
- 8 November 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (22) , 3589
Unsupervised Voting for Detecting the Algorithmic Solving Strategy in Competitive Programming Solutions
The problem of source-code analysis using machine-learning techniques has gained much attention recently, as several powerful code-embedding methods have been created. Having different embedding methods available for source code has opened the way to tackling many practical problems in source-code analysis. This paper addresses the problem of determining the number of distinct algorithmic strategies that may be found in a set of correct solutions to a competitive programming problem. To achieve this, we employ a novel unsupervised algorithm that uses a multiview interpretation of data based on different embedding and clustering methods, a multidimensional assignment problem (MAP) to determine a subset of a higher probability of correctness, and a self-training method based on voting to determine the correct clusters of the remaining set. We investigate the following two aspects: (1) whether the proposed unsupervised approach outperforms existing methods when the number K of distinct algorithmic strategies is known and (2) Whether the approach can also be applied to determine the optimal value of K. We have addressed these using seven embedding methods with three clustering strategies in a data-analysis pipeline that tackles the previously described issues on a newly created dataset consisting of 15 algorithmic problems. According to the results, for the first aspect, the proposed unsupervised voting algorithm significantly improves the baseline clustering approach for a known K. This improvement was observed across all problems in the dataset, except one. In the case of the second one, we prove that the proposed method has a negative impact on determining the optimal number K. Scale-up of the data-analysis pipeline to datasets of thousands of problems may yield the ability to profoundly understand and learn about the innovative process of correctly designing and writing code in the context of competitive programming or even industry code.
- 8 November 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (22) , 3589
Open AccessUnsupervised Voting for Detecting the Algorithmic Solving Strategy in Competitive Programming Solutions
The problem of source-code analysis using machine-learning techniques has gained much attention recently, as several powerful code-embedding methods have been created. Having different embedding methods available for source code has opened the way to tackling many practical problems in source-code analysis. This paper addresses the problem of determining the number of distinct algorithmic strategies that may be found in a set of correct solutions to a competitive programming problem. To achieve this, we employ a novel unsupervised algorithm that uses a multiview interpretation of data based on different embedding and clustering methods, a multidimensional assignment problem (MAP) to determine a subset of a higher probability of correctness, and a self-training method based on voting to determine the correct clusters of the remaining set. We investigate the following two aspects: (1) whether the proposed unsupervised approach outperforms existing methods when the number K of distinct algorithmic strategies is known and (2) Whether the approach can also be applied to determine the optimal value of K. We have addressed these using seven embedding methods with three clustering strategies in a data-analysis pipeline that tackles the previously described issues on a newly created dataset consisting of 15 algorithmic problems. According to the results, for the first aspect, the proposed unsupervised voting algorithm significantly improves the baseline clustering approach for a known K. This improvement was observed across all problems in the dataset, except one. In the case of the second one, we prove that the proposed method has a negative impact on determining the optimal number K. Scale-up of the data-analysis pipeline to datasets of thousands of problems may yield the ability to profoundly understand and learn about the innovative process of correctly designing and writing code in the context of competitive programming or even industry code.
- 25 June 2025
- Published by Springer Nature
- p. 141-155
Domain Adaptation for Automated Tag Prediction in Competitive Programming
- 25 June 2025
- Published by Springer Nature
- p. 141-155
Domain Adaptation for Automated Tag Prediction in Competitive Programming
- Open Access
- 23 June 2025
- Published by MDPI AG in Information
- Vol. 16 (7) , 524
GIT-CXR: End-to-End Transformer for Chest X-Ray Report Generation
Medical imaging is crucial for diagnosing, monitoring, and treating medical conditions. The medical reports of radiology images are the primary medium through which medical professionals can attest to their findings, but their writing is time-consuming and requires specialized clinical expertise. Therefore, the automated generation of radiography reports has the potential to improve and standardize patient care and significantly reduce the workload of clinicians. Through our work, we have designed and evaluated an end-to-end transformer-based method to generate accurate and factually complete radiology reports for X-ray images. Additionally, we are the first to introduce curriculum learning for end-to-end transformers in medical imaging and demonstrate its impact in obtaining improved performance. The experiments were conducted using the MIMIC-CXR-JPG database, the largest available chest X-ray dataset. The results obtained are comparable with the current state of the art on the natural language generation (NLG) metrics BLEU and ROUGE-L, while setting new state-of-the-art results on F1 examples-averaged F1-macro and F1-micro metrics for clinical accuracy and on the METEOR metric widely used for NLG.
- 23 June 2025
- Published by MDPI AG in Information
- Vol. 16 (7) , 524
Open AccessGIT-CXR: End-to-End Transformer for Chest X-Ray Report Generation
Medical imaging is crucial for diagnosing, monitoring, and treating medical conditions. The medical reports of radiology images are the primary medium through which medical professionals can attest to their findings, but their writing is time-consuming and requires specialized clinical expertise. Therefore, the automated generation of radiography reports has the potential to improve and standardize patient care and significantly reduce the workload of clinicians. Through our work, we have designed and evaluated an end-to-end transformer-based method to generate accurate and factually complete radiology reports for X-ray images. Additionally, we are the first to introduce curriculum learning for end-to-end transformers in medical imaging and demonstrate its impact in obtaining improved performance. The experiments were conducted using the MIMIC-CXR-JPG database, the largest available chest X-ray dataset. The results obtained are comparable with the current state of the art on the natural language generation (NLG) metrics BLEU and ROUGE-L, while setting new state-of-the-art results on F1 examples-averaged F1-macro and F1-micro metrics for clinical accuracy and on the METEOR metric widely used for NLG.
- Open Access
- 24 May 2025
- Published by MDPI AG in Information
- Vol. 16 (6) , 434
Multihead Average Pseudo-Margin Learning for Disaster Tweet Classification
During natural disasters, social media platforms, such as X (formerly Twitter), become a valuable source of real-time information, with eyewitnesses and affected individuals posting messages about the produced damage and the victims. Although this information can be used to streamline the intervention process of local authorities and to achieve a better distribution of available resources, manually annotating these messages is often infeasible due to time and cost constraints. To address this challenge, we explore the use of semi-supervised learning, a technique that leverages both labeled and unlabeled data, to enhance neural models for disaster tweet classification. Specifically, we investigate state-of-the-art semi-supervised learning models and focus on co-training, a less-explored approach in recent years. Moreover, we propose a novel hybrid co-training architecture, Multihead Average Pseudo-Margin, which obtains state-of-the-art results on several classification tasks. Our approach extends the advantages of the voting mechanism from Multihead Co-Training by using the Average Pseudo-Margin (APM) score to improve the quality of the pseudo-labels and self-adaptive confidence thresholds for improving imbalanced classification. Our method achieves up to 7.98% accuracy improvement in low-data scenarios and 2.84% improvement when using the entire labeled dataset, reaching 89.55% accuracy on the Humanitarian task and 91.23% on the Informative task. These results demonstrate the potential of our approach in addressing the critical need for automated disaster tweet classification. We made our code publicly available for future research.
- 24 May 2025
- Published by MDPI AG in Information
- Vol. 16 (6) , 434
Open AccessMultihead Average Pseudo-Margin Learning for Disaster Tweet Classification
During natural disasters, social media platforms, such as X (formerly Twitter), become a valuable source of real-time information, with eyewitnesses and affected individuals posting messages about the produced damage and the victims. Although this information can be used to streamline the intervention process of local authorities and to achieve a better distribution of available resources, manually annotating these messages is often infeasible due to time and cost constraints. To address this challenge, we explore the use of semi-supervised learning, a technique that leverages both labeled and unlabeled data, to enhance neural models for disaster tweet classification. Specifically, we investigate state-of-the-art semi-supervised learning models and focus on co-training, a less-explored approach in recent years. Moreover, we propose a novel hybrid co-training architecture, Multihead Average Pseudo-Margin, which obtains state-of-the-art results on several classification tasks. Our approach extends the advantages of the voting mechanism from Multihead Co-Training by using the Average Pseudo-Margin (APM) score to improve the quality of the pseudo-labels and self-adaptive confidence thresholds for improving imbalanced classification. Our method achieves up to 7.98% accuracy improvement in low-data scenarios and 2.84% improvement when using the entire labeled dataset, reaching 89.55% accuracy on the Humanitarian task and 91.23% on the Informative task. These results demonstrate the potential of our approach in addressing the critical need for automated disaster tweet classification. We made our code publicly available for future research.
- Open Access
Aegis2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails
As Large Language Models (LLMs) and generative AI become increasingly widespread, concerns about content safety have grown in parallel. Currently, there is a clear lack of high-quality, human-annotated datasets that address the full spectrum of LLM-related safety risks and are usable for commercial applications. To bridge this gap, we propose a comprehensive and adaptable taxonomy for categorizing safety risks, structured into 12 top-level hazard categories with an extension to 9 fine-grained subcategories. This taxonomy is designed to meet the diverse requirements of downstream users, offering more granular and flexible tools for managing various risk types. Using a hybrid data generation pipeline that combines human annotations with a multi-LLM "jury" system to assess the safety of responses, we obtain Aegis 2.0, a carefully curated collection of 34,248 samples of human-LLM interactions, annotated according to our proposed taxonomy. To validate its effectiveness, we demonstrate that several lightweight models, trained using parameter-efficient techniques on Aegis 2.0, achieve performance competitive with leading safety models fully fine-tuned on much larger, non-commercial datasets. In addition, we introduce a novel training blend that combines safety with topic following data.This approach enhances the adaptability of guard models, enabling them to generalize to new risk categories defined during inference. We plan to open-source Aegis 2.0 data and models to the research community to aid in the safety guardrailing of LLMs.
Open AccessAegis2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails
As Large Language Models (LLMs) and generative AI become increasingly widespread, concerns about content safety have grown in parallel. Currently, there is a clear lack of high-quality, human-annotated datasets that address the full spectrum of LLM-related safety risks and are usable for commercial applications. To bridge this gap, we propose a comprehensive and adaptable taxonomy for categorizing safety risks, structured into 12 top-level hazard categories with an extension to 9 fine-grained subcategories. This taxonomy is designed to meet the diverse requirements of downstream users, offering more granular and flexible tools for managing various risk types. Using a hybrid data generation pipeline that combines human annotations with a multi-LLM "jury" system to assess the safety of responses, we obtain Aegis 2.0, a carefully curated collection of 34,248 samples of human-LLM interactions, annotated according to our proposed taxonomy. To validate its effectiveness, we demonstrate that several lightweight models, trained using parameter-efficient techniques on Aegis 2.0, achieve performance competitive with leading safety models fully fine-tuned on much larger, non-commercial datasets. In addition, we introduce a novel training blend that combines safety with topic following data.This approach enhances the adaptability of guard models, enabling them to generalize to new risk categories defined during inference. We plan to open-source Aegis 2.0 data and models to the research community to aid in the safety guardrailing of LLMs.