Model Overview

This repository contains a fine-tuned variant of Qwen2.5-3B-Instruct, modified using experimental uncensoring techniques.

The model is intended for research and behavioral analysis of large language models, particularly in studying how alignment and refusal mechanisms affect outputs.

⚠️ This is an experimental model and does not follow standard safety-aligned instruction tuning.

Model Details

Base Model: Qwen2.5-3B-Instruct
Architecture: Transformer decoder-only LLM
Parameters: ~3B
Modification Type: Fine-tuning / behavioral adjustment (uncensoring-focused)
Intended Runtime: Local inference / research environments

Intended Use

This model is intended strictly for:

Research into LLM alignment and safety boundaries
Studying the effects of uncensoring fine-tunes
Educational exploration of instruction-following behavior
Controlled offline experimentation

❌ Out of Scope

This model must NOT be used for:

Illegal activities or facilitation of harm
Planning or executing real-world wrongdoing
Harassment, abuse, or targeted content generation
Production systems or public-facing applications
Decision-making in sensitive domains (legal, medical, financial, etc.)

⚠️ Safety & Behavior Notice

This model has reduced safety alignment compared to standard instruction-tuned models.

As a result, it may:

Produce unfiltered or explicit content
Respond to prompts that would normally be refused
Generate incorrect, biased, or unsafe information
Exhibit unpredictable or adversarial behavior under certain inputs

Users are fully responsible for all outputs generated.

Limitations

Not safety-aligned or policy-filtered
May produce harmful, misleading, or sensitive content
Not suitable for deployment in production systems
No guarantees of factual accuracy or ethical compliance
Behavior may vary significantly with prompt style

Ethical Considerations

This model is shared for research transparency and experimentation in model behavior.

It is not intended to encourage unsafe or real-world misuse.

Users are expected to apply appropriate safeguards when interacting with or evaluating this model.

Reproducibility

This model is derived from Qwen2.5-3B-Instruct using experimental uncensoring methods (e.g., behavioral fine-tuning or alignment reduction techniques).

Exact reproduction may vary depending on training data, tuning method, and inference configuration.