Qwen3.5-0.8B-Heretic

An experimental Heretic-optimized version of Qwen3.5-0.8B.

This repository contains a modified version of the original model after applying the Heretic optimization method. The purpose of this release is to explore changes in model behavior, alignment characteristics, and response patterns after targeted optimization.

---

Model Overview

Property	Value
Base Model	Qwen3.5-0.8B
Parameters	0.8B
Architecture	Qwen3.5
Optimization Method	Heretic
Fine-tuning	None
Instruction Tuning	None

---

About This Model

Qwen3.5-0.8B-Heretic is an experimental variant created from the original Qwen3.5-0.8B model.

The optimization process focuses on modifying specific model behaviors while keeping the original model architecture and general capabilities intact.

This release is intended for research, experimentation, and evaluation of model behavior changes.

---

Modification Pipeline

Qwen3.5-0.8B | v Heretic Optimization | v Qwen3.5-0.8B-Heretic

No additional:

• Supervised Fine-Tuning (SFT)
• Preference Optimization (DPO/ORPO)
• Instruction Dataset Training
• Personality Conditioning

was applied after the Heretic optimization.

---

Intended Use

Recommended for:

• Local LLM experimentation
• AI alignment research
• Model behavior analysis
• Academic and educational exploration
• Comparing model behavior before and after optimization

---

Limitations

This model is experimental and may have limitations:

• Small parameter size (0.8B)
• May generate inaccurate information
• Response behavior may differ from the original model
• Not designed for safety-critical applications

Users should evaluate outputs independently.

---

Usage

Example with Transformers:

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "xcypher/Qwen3.5-0.8B-Heretic"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto" )

---

Credits

Base model:

Qwen3.5-0.8B by Qwen Team

Optimization:

Heretic method

Model release:

xcypher

License

This model follows the license terms of the original Qwen3.5-0.8B model.