Researchers have introduced Lev-3B, a compact 3-billion parameter probabilistic decision model designed for log analysis tasks. This model, built upon Granite-4.1-3B-Base, utilizes a novel two-stage training recipe. The first stage involves learning bounded natural-language decisions without log data, followed by a log stage that mixes target tasks with general log data from 35 auxiliary tasks. This curriculum-based approach demonstrates superior transfer learning compared to training solely on log data, and the inclusion of general log data enhances performance on unseen sources. While Lev-3B outperforms prompted frontier models on anomaly detection at a lower serving cost, it lags in attack-technique classification and shows less reliable probability calibration on unfamiliar systems. AI
IMPACT Offers a more cost-effective solution for anomaly detection in log analysis compared to larger models.
RANK_REASON The cluster contains a research paper detailing a new model and its training methodology. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- Connected Papers
- DagsHub
- Gotit.pub
- Granite-4.1-3B-Base
- Hugging Face
- IArxiv
- Lev-3B
- Litmaps
- ScienceCast
- Scite
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
