tank-qwen3.5-4b-v0.3

Experimental QLoRA finetune of Qwen3.5-4B focused on ideological and argumentative steering.

This model was created as a learning and experimentation project to explore:

• semantic steering
• ontology shaping
• reasoning drift
• lightweight LoRA finetuning
• dataset mixing and curation

Base Model

• Qwen3.5-4B

Training Method

• QLoRA
• Unsloth
• 4-bit training

Datasets

Primary datasets used during experimentation:

• WokeAI/polititune-tankie-warmup-3
• WokeAI/tankie-special-prompts

Characteristics

The finetune intentionally biases the model toward:

• marxist/materialist analysis
• class-based framing
• ideological consistency
• argumentative persistence

The model may:

• strongly reinterpret political and economic questions through class analysis
• provide ideological or revolutionary framing
• present controversial political opinions confidently

This behavior is intentional and part of the experiment.

Limitations

This is an experimental finetune and was not trained for:

• factual neutrality
• balanced political discussion
• historical accuracy
• safety-critical usage

Outputs may contain:

• ideological bias
• historical simplifications
• overconfident reasoning
• persuasive argumentative structures

Intended Use

• research
• experimentation
• RP/persona exploration
• semantic steering studies
• LoRA experimentation

Not Intended For

• political persuasion
• misinformation
• real-world activism guidance
• factual/historical reference
• safety-critical systems

Notes

This project was created primarily to understand how small curated datasets can influence:

• reasoning style
• interpretative frameworks
• conversational behavior
• semantic priorities

License

Please refer to the license of the original base model and datasets.