Dingdust/ContextPilot-E4B-heretic
Dingdust/ContextPilot-E4B-heretic is a 7.9 billion parameter decensored version of the Tencent ContextPilot-E4B model, created using the Heretic v1.4.0 framework. This model is specifically designed for proactive context management in long-horizon language-model agents, teaching them to plan, maintain long-term memory, and offload less useful context. It excels in tasks requiring continuous reasoning and tool use, particularly in long-context question answering and deep search scenarios.
Loading preview...
Overview
Dingdust/ContextPilot-E4B-heretic is a 7.9 billion parameter model, a decensored variant of Tencent's ContextPilot-E4B, processed with Heretic v1.4.0. ContextPilot-E4B is a Gemma4-E4B checkpoint of the ContextPilot framework, which focuses on proactive context management for long-horizon language-model agents. It enables agents to plan, manage long-term memory, and intelligently offload less relevant context while performing reasoning and tool-use tasks.
Key Capabilities
- Proactive Context Management: Teaches agents to manage their working context efficiently over long interactions.
- Enhanced Toolset: Incorporates an extended context-management toolset including planning, structured memory, retrieval, and soft context offloading.
- Fine-grained RL: Utilizes context-aware partial rollout and fine-grained credit assignment to optimize context-editing decisions.
- Decensored Output: The "heretic" modification aims to reduce refusals, showing 32/100 refusals compared to 100/100 for the original model.
Good For
- Research on proactive context management and long-horizon agents.
- Applications requiring long-context question answering.
- Deep search tasks where agents need to iteratively retrieve and integrate dispersed information.
- Scenarios where a less restrictive model response is desired compared to the original ContextPilot-E4B.