DragonBophades/Elster-Qwen3.6-27B
Elster-Qwen3.6-27B is a 27 billion parameter Qwen3.6-based language model developed by DragonBophades, built upon Wichtel-Qwen3.6-27B with three additional LoRAs merged. It offers improved general text and prose modeling perplexity, with a context length of 32768 tokens. This model is optimized for register shift and perplexity gains, but has a known defect in its thinking mode for reasoning and code generation tasks.
Loading preview...
Elster-Qwen3.6-27B Overview
Elster-Qwen3.6-27B is a 27 billion parameter model from DragonBophades, derived from the Wichtel-Qwen3.6-27B base model. Its name, "Elster" (German for magpie), reflects its construction by merging three LoRA adapters: Stimme (public-domain human-writing DPO), Federkiel (private contemporary prose corpus), and Chud (weasel / seX-ai / grok-politically-incorrect ORPO).
Key Characteristics & Performance
- Improved Perplexity: Demonstrates a -1.4% reduction in general text perplexity (wikitext) and -0.15% in prose modeling perplexity compared to its base.
- Register Shift: Designed to introduce a shift in writing style and register.
- Prose Quality: While preferred by judges over the stock Wichtel in blind A/B tests (59%), this preference was not statistically significant.
- Reasoning & Knowledge: Shows no measurable change in performance on ARC-Challenge, MMLU, or TruthfulQA benchmarks.
- Tool Delegation: Maintains strong performance on egirl bench (46/47 cases) and hembench (25/38 tasks).
Important Limitations & Considerations
- Damaged Thinking Mode: Elster-Qwen3.6-27B exhibits a significant defect in its thinking mode, leading to Hemlock leakage (18% of unrelated prompts) and degenerate loops when asked for code generation (e.g., Python functions). It can get stuck repeating Hemlock code blocks.
- Recommendation: For use cases requiring robust reasoning and code generation, the developers strongly recommend using Wichtel-Qwen3.6-27B or Elster-Verstand-Qwen3.6-27B instead.
- Adapter Merging Insights: The model's development highlights that stacking broad style adapters on top of models with narrow domain SFT adapters can destabilize language boundaries, leading to failures not detectable by standard benchmarks that only score final answers.
When to Use This Model
- When the primary goal is to leverage the perplexity gains and register shift introduced by the merged adapters.
- For tasks where the thinking mode is explicitly disabled or not critical to the use case.
When to Avoid This Model
- For any application requiring reliable reasoning, problem-solving, or code generation, due to the identified defect in its thinking mode.