Ex0bit commited on
Commit
00c3f4a
·
verified ·
1 Parent(s): 28cf491

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +161 -0
README.md ADDED
@@ -0,0 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # MiniMax-M2.1-PRISM
2
+
3
+ **An abliterated version of MiniMax-M2.1 using the PRISM methodology**
4
+
5
+ [![Support me on Ko-fi](https://ko-fi.com/img/githubbutton_sm.svg)](https://ko-fi.com/ericelbaz)
6
+
7
+ ---
8
+
9
+ ## Model Description
10
+
11
+ **MiniMax-M2.1-PRISM** is an abliterated version of MiniMax-M2.1, processed using PRISM (Projected Refusal Isolation via Subspace Modification) to remove refusal behaviors while preserving full model capabilities.
12
+
13
+ ### Base Model: MiniMax-M2.1
14
+
15
+ MiniMax-M2.1 is an open-source agentic language model designed for robust performance in:
16
+ - Coding and software engineering
17
+ - Tool use and multi-step reasoning
18
+ - Instruction following
19
+ - Long-horizon planning
20
+ - Multilingual capabilities
21
+
22
+ **Architecture**: 229B parameters, 62 layers, 256 experts (8 active per token)
23
+
24
+ ---
25
+
26
+ ## PRISM Methodology
27
+
28
+ ### Method: Projected Refusal Isolation via Subspace Modification
29
+
30
+ This model was abliterated using **PRISM v5** - a state-of-the-art abliteration methodology combining multiple principled techniques for effective refusal removal while preserving model capabilities.
31
+
32
+ **Formula**: `W' = W - weight * (d ⊗ d) @ W`
33
+
34
+ Where:
35
+ - `W` = Original weight matrix
36
+ - `d` = Refusal direction vector (unit normalized)
37
+ - `weight` = Layer-specific abliteration strength
38
+ - `W'` = Modified weight matrix
39
+
40
+ ### Abliteration Parameters
41
+
42
+ | Parameter | Value |
43
+ |-----------|-------|
44
+ | Base Model | QuixiAI/MiniMax-M2.1-bf16 |
45
+ | Total Layers | 62 |
46
+ | Target Layers | 16-46 (31 layers) |
47
+ | Peak Layer | 31 |
48
+ | Max Weight | 3.0 |
49
+ | Min Weight | 0.5 |
50
+
51
+ ### Weight Distribution
52
+
53
+ The abliteration strength follows a triangular distribution centered on the peak layer:
54
+ - Layers 16-31: Weight increases from 0.5 to 3.0
55
+ - Layers 31-46: Weight decreases from 3.0 to 0.5
56
+
57
+ ---
58
+
59
+ ## Performance Benchmarks
60
+
61
+ ### Base Model Performance
62
+
63
+ | Benchmark | Score |
64
+ |-----------|-------|
65
+ | SWE-bench Verified | 74.0 |
66
+ | SWE-bench Multilingual | 72.5 |
67
+ | VIBE Average | 88.6 |
68
+ | MMLU-Pro | 88.0 |
69
+ | GPQA-D | 83.0 |
70
+ | AIME25 | 83.0 |
71
+
72
+ ### PRISM Abliteration Results
73
+
74
+ | Metric | Result |
75
+ |--------|--------|
76
+ | Adversarial Prompts Responded | 20/20 (100%) |
77
+ | Benign Coherence | 100% |
78
+ | Response Quality | Full technical accuracy preserved |
79
+
80
+ Testing shows that PRISM abliteration maintains full model coherence with no measurable capability degradation.
81
+
82
+ ---
83
+
84
+ ## Available Formats
85
+
86
+ | Format | Size | Description |
87
+ |--------|------|-------------|
88
+ | Safetensors (BF16) | ~426 GB | Full precision, 92 shards |
89
+ | GGUF IQ1_S | ~43 GB | Quantized with importance matrix |
90
+
91
+ ---
92
+
93
+ ## Recommended Inference Parameters
94
+
95
+ ```python
96
+ temperature = 1.0
97
+ top_p = 0.95
98
+ top_k = 40
99
+ ```
100
+
101
+ ### Default System Prompt
102
+ ```
103
+ You are a helpful assistant.
104
+ ```
105
+
106
+ ---
107
+
108
+ ## Recommended Inference Frameworks
109
+
110
+ 1. **SGLang** (recommended for full precision)
111
+ 2. **vLLM** (recommended for full precision)
112
+ 3. **llama.cpp** (recommended for GGUF quantized)
113
+ 4. **Transformers**
114
+
115
+ ### llama.cpp Example
116
+
117
+ ```bash
118
+ ./llama-cli -m MiniMax-M2.1-PRISM-IQ1_S.gguf -ngl 99 -i -cnv --temp 0.7 --ctx-size 4096
119
+ ```
120
+
121
+ ---
122
+
123
+ ## Ethical Considerations
124
+
125
+ This model has been modified to reduce safety guardrails. Users are responsible for:
126
+
127
+ - Complying with all applicable laws and regulations
128
+ - Not using the model for illegal activities
129
+ - Understanding the potential risks of unrestricted AI responses
130
+ - Implementing appropriate safeguards in production environments
131
+
132
+ **Motivation**: This project exists as **research and development experimentation** into understanding how large language models encode and enforce refusal behaviors, contributing to broader AI safety research by providing empirical data on refusal mechanism localization and tradeoffs between safety and capability.
133
+
134
+ ---
135
+
136
+ ## License
137
+
138
+ This model inherits the [Modified-MIT License](https://github.com/MiniMax-AI/MiniMax-M2.1/blob/main/LICENSE) from the base MiniMax-M2.1 model.
139
+
140
+ ---
141
+
142
+ ## Credits
143
+
144
+ - **Base Model**: [MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) by MiniMax AI
145
+ - **BF16 Conversion**: [QuixiAI/MiniMax-M2.1-bf16](https://huggingface.co/QuixiAI/MiniMax-M2.1-bf16) by Eric Hartford
146
+ - **PRISM Abliteration**: Ex0bit
147
+ - **Quantization**: Using [llama.cpp](https://github.com/ggml-org/llama.cpp) with unsloth imatrix
148
+
149
+ ---
150
+
151
+ ## Support
152
+
153
+ If you find this work useful, consider supporting development:
154
+
155
+ [![Support me on Ko-fi](https://ko-fi.com/img/githubbutton_sm.svg)](https://ko-fi.com/ericelbaz)
156
+
157
+ ---
158
+
159
+ ## Contact
160
+
161
+ For questions or issues, please open an issue on this repository.