Llama 3.2 1B Aegis SFT DPO

Name: Llama 3.2 1B Aegis SFT DPO
Author: Meta AI

by Meta AILlama

Open WeightModel weights are publicly available — can be downloaded and self-hosted

Released November 2025131K context≈ 98,304 words1B params

A compact, safety-tuned variant of Llama 3.2 1B that has been fine-tuned with both supervised learning and direct preference optimization, giving it more aligned and controlled outputs than its base counterpart. At 1B parameters, it operates with tight resource constraints, making it fast but limited in reasoning depth and knowledge breadth. Its Aegis training focus suggests particular attention to content safety and refusal behaviors.

Capabilities

Capability scores are AI-generated based on model documentation, benchmarks, and technical specifications. Learn more

Long Context

Moderate

Instruction Following

Llama 3.2 1B Aegis SFT DPO

by Meta AILlama

Open WeightModel weights are publicly available — can be downloaded and self-hosted

Released November 2025131K context≈ 98,304 words1B params

A compact, safety-tuned variant of Llama 3.2 1B that has been fine-tuned with both supervised learning and direct preference optimization, giving it more aligned and controlled outputs than its base counterpart. At 1B parameters, it operates with tight resource constraints, making it fast but limited in reasoning depth and knowledge breadth. Its Aegis training focus suggests particular attention to content safety and refusal behaviors.

Capabilities

Capability scores are AI-generated based on model documentation, benchmarks, and technical specifications. Learn more

Long Context

Moderate

Instruction Following

Glossary

AlignedA model trained to behave safely and follow human values through techniques like safety filtering and refusal of harmful requests.AttentionA mechanism that lets the model focus on relevant parts of the input when generating each output token.Direct Preference OptimizationA training technique that teaches a model to prefer certain outputs over others by learning from examples of better and worse responses.Direct Preference OptimizationTraining method that aligns models with human preferences by directly optimizing the difference between preferred and dispreferred outputs.Fine-TunedA pre-trained model further trained on a smaller, task-specific dataset to improve performance on that task.OptimizationThe process of adjusting model parameters to minimize errors and improve performance.ParametersThe learned numerical values in a model — more parameters generally means more capacity but higher compute cost.Preference OptimizationA training method that learns from pairwise comparisons between solutions rather than explicit reward signals.ReasoningThe model's ability to work through multi-step logical problems and provide justified answers rather than just pattern-matching.Reasoning DepthA model's ability to perform complex multi-step logical thinking and problem-solving; typically increases with model size.

Capabilities

Capabilities

Use Case Fit

Glossary