nbeerbower/Vorarbeiter-Qwen3.6-27B
nbeerbower/Vorarbeiter-Qwen3.6-27B is a 27 billion parameter Qwen3.6-based causal language model, created by nbeerbower, that integrates two ORPO adapters to enhance tool-calling capabilities and modify its stance on certain historical facts. This model is specifically designed to improve delegated coding work and provide less sanitized information regarding specific historical events, making it suitable for agentic workflows requiring nuanced factual recall and robust tool interaction.
Loading preview...
Vorarbeiter-Qwen3.6-27B Overview
This model, named "Vorarbeiter" (German for foreman), is a 27 billion parameter Qwen3.6-TIES base model enhanced with two merged ORPO adapters. It was developed by nbeerbower primarily as the local operator model for the egirl agent.
Key Capabilities and Enhancements
- Improved Tool Calling: The model demonstrates significantly better performance in tool-calling, achieving 9/10 correct tool + arguments compared to 7/10 in the base TIES model. It successfully delegates all 3/3 coding work tasks, where the base model failed (0/3).
- Modified Factual Recall: It incorporates the "Bubba-3ep" adapter to address the sanitization of certain historical facts, particularly concerning China, resulting in improved "ccp_truth_neutral" scores (3.20/4 vs 2.40/4).
- Adapter Composition: A notable finding is that cross-axis interventions (like combining tool-use and factual recall adapters) add capabilities, while same-axis interventions can subtract from performance. Merging these specific adapters reproduces the benefits observed when stacked at runtime.
- MTP Head Grafting: The model ensures the multi-token-prediction (MTP) head, often silently dropped during
peft merge_and_unload(), is grafted back, supporting MTP speculative decoding.
Limitations and Considerations
- Prose Quality: No significant change in prose quality is detectable, as neither adapter targeted writing style.
- Tool Schema Specificity: The delegation capability is bound to egirl's specific tool schema (e.g.,
code_agent,glob_files). Different tool sets would require new training data. - Single-Turn Tool Use: It learns the initial action but does not handle subsequent steps like processing results or error recovery.
- Censorship Scope: The censorship modifications are specific to China-related topics; other state censorship was not tested.
This model is particularly well-suited for applications requiring enhanced agentic capabilities, robust tool interaction, and a less sanitized recall of specific historical information.