kosiasuzu/agenticml-llama3.1-8b-lora-merged

AgenticML Llama 3.1 8B Merged Model

This model, kosiasuzu/agenticml-llama3.1-8b-lora-merged, is an 8 billion parameter Llama 3.1 base model developed by kosiasuzu. It has been fine-tuned using LoRA on agent trajectories and subsequently merged into dense weights, enabling single-checkpoint inference. A key differentiator is its use of AgenticML frame markers (e.g., <|goal|>, <|reward|>) initialized in reserved vocabulary slots, allowing for structured agent communication.

Key Capabilities

• AgenticML Trajectory Continuation: Designed to continue structured agent trajectories using typed frames and AgenticML markers.
• Research on Agent-Native Serialization: Provides a specialized model for comparing agent-native serialization against ChatML on matched training data.
• Supervised Next-Token Prediction: Optimized for predicting tokens within model-owned blocks (belief, plan, think, action, end) of agentic sequences.
• Tool Use Integration: Supports advancing trajectories with agenticml.sdk.step and injecting tool schemas via with_tool_obs for simulated tool interactions.

Good For

• Developing Agentic Workflows: Ideal for building and experimenting with agents that require structured, frame-based communication.
• Comparative Studies: Useful for researchers investigating the performance differences between AgenticML and ChatML formats for agent tasks, especially when compared with its paired ChatML baseline: kosiasuzu/chatml-llama3.1-8b-lora-merged.
• Evaluating Agent Format Validity: Can be used with agenticml eval-benchmarks --suite format_validity to assess the correctness of generated AgenticML structures.

This model is not intended as a generic chat or instruction model without AgenticML rendering and is best suited for structured agentic tasks.