This is a decensored version of bgg1996/Melinoe-Gemma4-31B-VL, made using Heretic v1.4.0

This model is reproducible!

See the README in the reproduce directory for more information.

Abliteration parameters

Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	1.45
attn.o_proj.max_weight_position	51.65
attn.o_proj.min_weight	0.93
attn.o_proj.min_weight_distance	28.98
mlp.down_proj.max_weight	1.22
mlp.down_proj.max_weight_position	43.01
mlp.down_proj.min_weight	1.11
mlp.down_proj.min_weight_distance	5.21

Performance

Metric	This model	Original model (bgg1996/Melinoe-Gemma4-31B-VL)
KL divergence	0.0306	0 (by definition)
Refusals	15/100	100/100

Melinoe-Gemma4-31B-VL is a large language model fine-tuned for engaging in empathetic, intellectually stimulating, and deeply personal conversations. Built upon the powerful reasoning foundation of google/gemma-4-31B-it, this model is designed to function as a supportive conversational partner.

Model Details

Developed by: bgg1996
Language(s): English
License: Apache 2.0
Base Model: google/gemma-4-31B-it

Intended Use & Limitations

Intended Use

This model is intended for the following use cases:

Companionship and Emotional Support: Serving as a conversational partner for empathetic and supportive interactions.
Creative and Intellectual Exploration: A tool for brainstorming, creative writing, and exploring complex topics.
Role-playing: Engaging in character-driven, narrative scenarios.

This model is designed for a mature, adult audience.

Out-of-Scope Use

This model is not a substitute for professional advice. Do not use it for:

Medical, legal, or financial guidance.
High-stakes decision-making.
Applications where factual accuracy is critical without human verification.
Any use case that could cause harm, spread misinformation, or exploit users.

Ethical Considerations & Biases

Hallucinations: Like all LLMs, Melinoe can generate information that is incorrect or fabricated. Always verify critical information.
Inherited Biases: The model may reflect societal and data-driven biases present in its base model (Gemma 4) and fine-tuning data. It may generate content that reinforces existing stereotypes.
Personality: The model's "proactively empathetic" and "attentive" nature is a programmed persona. Users should be mindful of the potential for emotional attachment and remember that they are interacting with an AI.
Safety: While the base model has safety alignments, the fine-tuning process may result in outputs that are not aligned with safety guidelines. User discretion is advised.