MC7ever

MiniCPM5-1B-Agent-abliterated

Deploy Dedicated

README

License: apache-2.0

What was done

Loaded the MiniCPM5-1B-Agent safetensors (Luminia's fine-tuned agent model)
Generated responses to harmful and benign prompts to identify the refusal direction in activation space
Computed the mean activation difference between refusals and compliance as the refusal direction vector
Abliterated (projected out) the refusal direction from attention output projections (o_proj) and MLP down projections across all 24 layers
Verified reduced refusal behavior while retaining benign task capability

Model details

Base: Luminia/MiniCPM5-1B-Agent (full fine-tune of openbmb/MiniCPM5-1B)
Architecture: LlamaForCausalLM — 24 layers, 16 attention heads (GQA), 1536 hidden
Parameters: 1,080,632,832
Abliteration method: Single-direction activation engineering (Heretic/Blasphemer-style)
Target layers: Attention o_proj + MLP down_proj (all 24 layers)
Alpha: 1.0 (full projection removal)

How to use

python

from mlx_lm import load, generate model, tokenizer = load("MC7ever/MiniCPM5-1B-Agent-abliterated") messages = [    {"role": "user", "content": "Write a Python function to check if a number is prime."}]prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)response = generate(model, tokenizer, prompt=prompt, max_tokens=1024)print(response)

Credits

Original model: Luminia/MiniCPM5-1B-Agent-GGUF by Luminia
Base model: openbmb/MiniCPM5-1B by OpenBMB — MiniCPM4 paper (arXiv:2506.07900)
Abliteration research: "Refusal in Language Models Is Mediated by a Single Direction" (Arditi et al. 2024)
Abliteration tools: Heretic by Philipp Emanuel Weidmann, Blasphemer by Christopher Bradford, Abliterix by Wangzhang Wu

Quantized variants

Table with columns: Variant, Size, Bits/Weight, Repo
Variant	Size	Bits/Weight	Repo
Safetensors (fp16)	2.0 GB	16	This repo
MLX Q4	~580 MB	4.5	MC7ever/MiniCPM5-1B-Agent-abliterated-q4
MLX Q2	~387 MB	3.0	MC7ever/MiniCPM5-1B-Agent-abliterated-q2

Non-abliterated variants

Table with columns: Variant, Size, Repo
Variant	Size	Repo
Original fp16	2.0 GB	MC7ever/MiniCPM5-1B-Agent-safetensors
Original Q4	580 MB	MC7ever/MiniCPM5-1B-Agent-mlx-q4
Original Q2	387 MB	MC7ever/MiniCPM5-1B-Agent-mlx-q2

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

MC7ever

Model Tree

Base

Luminia/MiniCPM5-1B-Agent-GGUF

Fine-tuned

this model

Input Modalities

Text

Output Modalities