DragonBophades/Eintopf-Qwen3.8-27B
Eintopf-Qwen3.8-27B is a 27 billion parameter language model by DragonBophades, based on Qwen3.8, with nine LoRA adapters merged at full strength. This model primarily serves as a research artifact demonstrating that stacking multiple LoRAs does not necessarily improve performance over a single adapter, particularly for behavioral policies. It shows a modest improvement in multiple-choice reasoning but no significant change in tool-use behavior compared to its base model.
Loading preview...
Overview
Eintopf-Qwen3.8-27B is a 27 billion parameter model developed by DragonBophades, built upon the Qwen3.8-27B base. Its unique characteristic is the integration of nine distinct LoRA adapters (trained on the Qwen3.6 line) merged at full strength. This model is presented as a negative research result, highlighting that combining multiple adapters does not inherently lead to cumulative gains in all areas.
Key Findings & Capabilities
- Reasoning Improvement: Demonstrates a statistically significant, albeit modest, improvement in multiple-choice reasoning tasks (ARC-Challenge), scoring 57.86% compared to the base model's 52.51%. However, this gain is comparable to using just one specific adapter (
Lehrling) alone. - Behavioral Invariance: Crucially, the model shows no significant change in behavioral policies, such as tool selection or delegation, despite including adapters specifically designed for these tasks. This indicates that task vectors trained on one pretrain (Qwen3.6) may transfer for reasoning shifts but not for behavioral policies on a different base (Qwen3.8).
- No Hallucination: Maintains the base model's strong performance in not hallucinating tools, achieving 47/47 cases without inventing tools outside the provided set.
- Architectural Divergence: Analysis reveals that Qwen3.8 and Qwen3.6 differ most significantly in their MLPs, which is where most of the merged LoRA changes land, potentially explaining the split results between reasoning and behavior.
When to Use This Model
This model is primarily a research artifact for understanding the effects of LoRA merging and transferability across different base models. It is not recommended for general use cases where optimal performance is desired. For a more performant model, the README explicitly suggests using Wichtel-Qwen3.6-27B, which significantly outperforms Eintopf in both reasoning and behavioral benchmarks.