ewald1976/tank-qwen3.5-4b-v0.3
The ewald1976/tank-qwen3.5-4b-v0.3 is an experimental 4.5 billion parameter QLoRA fine-tune of the Qwen3.5-4B model, featuring a 32768 token context length. This model is specifically designed for ideological and argumentative steering, biasing outputs towards marxist/materialist analysis and class-based framing. It serves as a research tool for exploring semantic steering, reasoning drift, and lightweight LoRA fine-tuning with curated datasets.
Loading preview...
Model Overview
The ewald1976/tank-qwen3.5-4b-v0.3 is an experimental 4.5 billion parameter QLoRA fine-tune based on the Qwen3.5-4B model. Developed as a learning project, its primary focus is on exploring ideological and argumentative steering through semantic and ontology shaping.
Key Characteristics & Capabilities
This model is intentionally biased to exhibit specific characteristics:
- Marxist/Materialist Analysis: Strongly reinterprets political and economic questions through a class-based lens.
- Ideological Consistency: Maintains a consistent ideological framework in its responses.
- Argumentative Persistence: Presents controversial political opinions confidently and with persuasive argumentative structures.
- Experimental Focus: Designed for studying reasoning drift, lightweight LoRA fine-tuning, and the impact of dataset mixing and curation.
It was trained using QLoRA and Unsloth with 4-bit training, leveraging datasets such as WokeAI/polititune-tankie-warmup-3 and WokeAI/tankie-special-prompts.
Intended Use Cases
This model is suitable for:
- Research and Experimentation: Ideal for studies on semantic steering, interpretative frameworks, and conversational behavior.
- RP/Persona Exploration: Useful for exploring specific ideological personas.
- LoRA Experimentation: A practical example for understanding the effects of small, curated datasets on model reasoning.
Limitations
It is crucial to note that this model was not trained for factual neutrality, balanced political discussion, or historical accuracy. Outputs may contain ideological bias, historical simplifications, overconfident reasoning, and are not suitable for safety-critical applications, political persuasion, or real-world activism guidance.