audreyt/Gemma-3-TAIDE-12b-Chat-2602-heretic

Gemma-3-TAIDE-12b-Chat-2602 Abliterated Model

⚠️ This is a modified derivative model. It is NOT a official TAIDE model.

Model Description

This model is a derivative of the Gemma-3-TAIDE-12b-Chat-2602, developed by the National Applied Research Laboratories (財團法人國家實驗研究院, 國研院) under the TAIDE project (taide.tw).

The modification was performed using Heretic, an abliteration tool that adjusts model weights to reduce refusal behavior. This process modifies the model's internal representations to decrease the likelihood of refusing to respond to user prompts.

Modification Details

• Base Model: Gemma-3-TAIDE-12b-Chat-2602
• Modification Method: Abliteration via Heretic. Refusals: 10/100, KL Divergence: 0.0796
• What Changed: The model's refusal behavior has been significantly reduced. This was achieved by identifying and adjusting the directions in the model's residual stream that are associated with refusal outputs.
• What Did NOT Change: The model's general language capabilities, knowledge, and cultural context remain largely intact (as measured by KL divergence from the original distribution).

License & Attribution

TAIDE License

This derivative model is subject to the TAIDE模型授權條款 (TAIDE Model License Terms). A copy of the license is available at taide.tw.

Key obligations under the TAIDE license:

• This model originates from 國研院 TAIDE G類模型 — https://taide.tw
• Modifications have been clearly disclosed in this document (as required by Section 4(2))
• This derivative model must not be used for military or illegal purposes (Section 4(1))
• This derivative model must not be used to generate illegal or inappropriate content, including but not limited to: content that infringes on the rights of others, incites or promotes violence, misleads or promotes disinformation, or pornographic content (Section 4(1))
• Users and downstream recipients must comply with the full TAIDE license terms

Gemma Terms of Use

The base model is built upon Google's Gemma. Accordingly:

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

Users must also comply with the Gemma Prohibited Use Policy referenced therein.

Intended Use

This model is intended for research, experimentation, and applications where reduced refusal behavior is desired while maintaining the TAIDE model's strengths in Taiwanese sociocultural understanding.

Prohibited Use

Per the TAIDE license and Gemma Terms of Use, this model must not be used for:

• Military purposes
• Any illegal activity
• Generating content that infringes on the rights of others
• Inciting or promoting violence
• Creating or spreading misinformation
• Generating pornographic content
• Misrepresenting AI-generated content as human-generated
• Any use prohibited under the Gemma Prohibited Use Policy

Disclaimer

This is a community modification and is not endorsed by, affiliated with, or supported by 國研院 (NARLabs), the TAIDE project team, or Google. The creator(s) of this derivative model do not claim ownership of the TAIDE G類模型.

The removal of refusal behavior does not grant permission to generate prohibited content. Users remain fully responsible for ensuring their use complies with all applicable license terms and laws.

---

This README was generated to fulfill the disclosure and attribution requirements of the TAIDE Model License Terms (Section 4(2), 4(4), 4(5)) and the Gemma Terms of Use.