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

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 18, 2026License:llama3.1Architecture:Transformer Cold

The kosiasuzu/chatml-llama3.1-8b-lora-merged model is an 8 billion parameter Llama 3.1 causal language model, developed by the AgenticML project. It is LoRA-fine-tuned with ChatML templates and tool-call tokens, then merged into full weights for single-checkpoint inference. This model is specifically designed for research comparison in multi-turn ChatML agent trajectories, particularly against AgenticML-format models, and is not intended as a general-purpose chat product. It excels at continuing agentic conversations involving system, user, and tool messages, and assistant generations with tool payloads.

Loading preview...

Model Overview

This model, kosiasuzu/chatml-llama3.1-8b-lora-merged, is an 8 billion parameter Llama 3.1 causal language model developed by the AgenticML project. It has been fine-tuned using LoRA with ChatML templates and tool-call tokens, then merged into a single checkpoint for inference. It serves as a direct comparison arm for the AgenticML agent-format study, utilizing the same training data and recipe as its AgenticML-format counterpart, differing primarily in serialization.

Key Capabilities

  • ChatML Agent Trajectories: Designed to continue multi-turn agent conversations, handling system, user, and tool messages, as well as assistant generations with <|eot_id|> and tool payloads.
  • Tool-Call Integration: Incorporates tool-call tokens and simulated tool calls within its training, enabling structured interactions.
  • Research Comparison: Specifically intended for research to compare performance against AgenticML-format models on identical tasks and evaluation splits.
  • Llama 3.1 Base: Inherits the architecture and base capabilities of Meta's Llama 3.1 8B model.

Training and Data

The model was fine-tuned on the kosiasuzu/agenticml-agent-trajectory-dataset, focusing on the messages column (ChatML conversation derived from AgenticML trajectories). Training involved supervised next-token prediction with labels masked to assistant spans only. It was initialized from kosiasuzu/chatml-agent-llama-3.1-8b-init, which includes ChatML special-token rows.

Intended Use Cases

  • Agentic Research: Ideal for researchers studying agentic behavior and comparing different agent serialization formats (ChatML vs. AgenticML).
  • Tool-Use Evaluation: Useful for evaluating models' ability to process and generate structured tool calls within a ChatML context.
  • Format Validity Benchmarking: Can be used with the AgenticML CLI to run format-validity benchmarks.

Limitations and Considerations

  • Research-Specific: Not validated for general-purpose consumer chat or high-stakes applications.
  • Synthetic Data Biases: Trained largely on synthetic agent trajectories, which may introduce biases and domain limitations.
  • Simulated Tools: Tool calls in training data are simulated; real API integration requires external schemas and safety layers.

Finetunes
2 models