renaudb1999/le-harnais-ft-agentworld-1b
The renaudb1999/le-harnais-ft-agentworld-1b is a 1 billion parameter world-model student distilled from Qwen-AgentWorld-35B, built upon the Llama-3.2-1B-Instruct base model. This model is specifically designed as a Mac-runnable tool-outcome predictor, excelling in agentic environments. It achieves a token-F1 score of 0.826 and an OBSERVATION hit-rate of 60% against its teacher model, making it suitable for predicting outcomes in agent-based systems.
Loading preview...
Overview
renaudb1999/le-harnais-ft-agentworld-1b is a compact 1 billion parameter language model, distilled from the larger Qwen-AgentWorld-35B teacher model. It is built on the meta-llama/Llama-3.2-1B-Instruct architecture, making it part of the Llama family and subject to its community license. This model is specifically engineered to function as a world-model student and a tool-outcome predictor within agentic systems.
Key Capabilities
- Agentic World Modeling: Distilled to understand and predict outcomes in agent-based environments.
- Tool-Outcome Prediction: Specialized in forecasting the results of tool usage within an agent's operational context.
- Performance: Achieves a token-F1 score of 0.826 and an OBSERVATION hit-rate of 60% when compared to its Qwen-AgentWorld-35B teacher on a 40-example test set.
- Accessibility: Provided in multiple formats including
*.safetensorsfortransformersand*.gguf(Q4_K_M, Q8_0) forollamaorllama.cpp, making it Mac-friendly.
Good For
- Developers working on agent orchestration where predicting tool outcomes is crucial.
- Applications requiring a lightweight, yet capable, model for simulating agent interactions or understanding their operational flow.
- Environments where resource efficiency and Mac compatibility are important considerations for deploying agentic models.