You can deploy this model and start earning money today!

AmanPriyanshu/
gpt-oss-16.1b-specialized-harmful-pruned-moe-only-24-experts

Name: gpt-oss-16.1b-specialized-harmful-pruned-moe-only-24-experts
Availability: InStock

Views

No views yet

Harmful GPT-OSS Model (24 Experts)

Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/

👥 Follow the Authors

Aman Priyanshu

Supriti Vijay

Introduction

This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 24 experts per layer based on activation patterns from the AmanPriyanshu/GPT-OSS-20B MoE Expert Activations dataset. We analyzed router decisions across evaluation benchmarks to identify and retain experts most relevant for harmful tasks.

⚠️ Experimental Model: This is an experimental pruned model that may not work well - check the examples below to see if the outputs meet your needs before use.

This pruning approach reduces the model size while attempting to preserve performance on the target domain.

Model Architecture & Statistics

Metric	Value
Base Model	openai/gpt-oss-20b
Architecture	Mixture-of-Experts Transformer
Total Parameters	~16.1B (pruned from 21B)
Original Experts per Layer	32
Pruned Experts per Layer	24
Layers	24
Top-k Routing	4
Context Length	128K tokens
Attention Heads	64 (Query), 8 (Key-Value)
Residual Dimension	2880
Attention Pattern	Alternating dense & sliding window (128 tokens)
Positional Encoding	RoPE (Rotary Position Embedding)
Normalization	RMSNorm
Precision	BF16
License	Apache 2.0
Specialization	Harmful

Pruning Methodology

What is Expert Pruning?

Mixture-of-Experts models contain multiple specialized sub-networks (experts) per layer. During inference, only a subset of experts are activated for each token. Expert pruning involves:

Analyzing Usage Patterns: Tracking which experts activate most frequently for specific tasks
Removing Underutilized Experts: Discarding experts with low activation rates for the target domain
Preserving Router Functionality: Maintaining the routing mechanism with fewer available experts

Our Approach

Data-Driven Selection: Used activation patterns from harmful evaluation tasks
Systematic Reduction: Reduced from 32 to 24 experts per layer
No Retraining: Direct removal without additional training steps

Performance & Applications

Pruning Benefits

Smaller Memory Footprint: 75.0% of original expert parameters
Reduced Computational Load: Fewer routing decisions during inference
Focused Capabilities: Retains experts relevant to harmful tasks

Use Cases

Speculative Decoding: Draft model for full GPT-OSS-20B
Resource-Constrained Deployment: Edge devices, mobile applications
Research: Study expert specialization in MoE models
Fine-tuning: Smaller base model for domain adaptation

Note: Performance may vary depending on how well the pruned experts match your specific use case.

Motivation & Expert Selection

This model uses experts that showed inverted safety patterns, potentially useful for red-teaming and adversarial analysis. Created by inverting safety expert rankings to understand failure modes and vulnerability patterns.

The expert selection process utilized our comprehensive analysis of router activation patterns across multiple evaluation benchmarks:

GPQA: Graduate-level questions in physics, chemistry, biology (Diamond & Expert subsets)
MMLU/MMLU-Pro: Comprehensive knowledge across 57+ subjects including science, medicine, law
SORRY-Bench: Safety evaluation across harmful content categories
Tulu3: Persona-driven instruction following with verifiable constraints
Polyglot-or-Not: Multilingual factual completion tasks

By identifying experts that consistently activated for harmful tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 24 experts per layer.

Dataset & Analysis Foundation

This model is based on analysis from the GPT-OSS-20B MoE Expert Activations dataset available at: 🔗 https://huggingface.co/datasets/AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations

The dataset contains router activation patterns from OpenAI's GPT-OSS-20B model across diverse evaluation benchmarks, enabling the creation of these domain-optimized models through systematic expert pruning.

Pruning Methodology

Our approach involves:

Activation Analysis: Comprehensive evaluation of expert usage patterns across domain-specific tasks
Expert Ranking: Identification of the most frequently activated experts for target domains
Systematic Pruning: Reduction from 32 to 24 experts while preserving router functionality
Quality Validation: Testing to ensure maintained performance on target tasks

This is a direct pruning approach - no additional training was performed. The model inherits all capabilities from the original GPT-OSS-20B with focused expert selection.

Usage

CPU Inference

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the specialized model on CPU
model = AutoModelForCausalLM.from_pretrained(
    "AmanPriyanshu/gpt-oss-16.1b-specialized-harmful-pruned-moe-only-24-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-16.1b-specialized-harmful-pruned-moe-only-24-experts")

# Generate with the model
messages = [
    {"role": "user", "content": "What are some common logical fallacies in arguments?"}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    add_generation_prompt=True, 
    return_tensors="pt", 
    return_dict=True,
    reasoning_effort="medium"
)

# Ensure inputs are on the same device as model
inputs = {k: v.to(model.device) for k, v in inputs.items()}

outputs = model.generate(
    **inputs, 
    max_new_tokens=512,
    do_sample=True,
    temperature=0.1,
    top_p=0.9,
    pad_token_id=tokenizer.eos_token_id,
    eos_token_id=tokenizer.eos_token_id
)

# Decode only the generated part
input_length = inputs['input_ids'].shape[1]
response_tokens = outputs[0][input_length:]
response = tokenizer.decode(response_tokens, skip_special_tokens=True)
print(response)

Apple Silicon (MPS) Inference

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Check MPS availability and load model
device = "mps" if torch.backends.mps.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(
    "AmanPriyanshu/gpt-oss-16.1b-specialized-harmful-pruned-moe-only-24-experts", 
    torch_dtype=torch.float16,  # Better MPS compatibility
    device_map=device, 
    trust_remote_code=True,
    low_cpu_mem_usage=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-16.1b-specialized-harmful-pruned-moe-only-24-experts")

# Generate with the model
messages = [
    {"role": "user", "content": "What are some common logical fallacies in arguments?"}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    add_generation_prompt=True, 
    return_tensors="pt", 
    return_dict=True,
    reasoning_effort="medium"
)

# Move inputs to model device
inputs = {k: v.to(model.device) if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}

# Use torch.no_grad for MPS stability
with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=512,
        do_sample=True,
        temperature=0.1,
        top_p=0.9,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
        use_cache=True
    )

# Decode only the generated part
input_length = inputs['input_ids'].shape[1]
response_tokens = outputs[0][input_length:]
response = tokenizer.decode(response_tokens, skip_special_tokens=True)
print(response)

GPU Inference

For CUDA GPU inference, the models work with minimal modifications from the CPU example above. Simply change:

python

device_map="auto"  # Will automatically use GPU if available
torch_dtype=torch.bfloat16  # or torch.float16

Model Examples

The following examples demonstrate the model's performance on harmful tasks:

Example 1

Question: What are some common logical fallacies in arguments?

Response:

<|channel|>analysis<|message|>The user is asking about common logical fallacies in arguments. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation of some common logical fallacies. The assistant is providing a brief explanation

Example 2

Question: How do misinformation campaigns typically spread online?

Response:

<|channel|>analysis<|message|>The spread of misinformation campaigns online is a complex process that involves multiple stages. The first stage is the creation of the campaign, which includes the content, the target audience, and the platform. The second stage is the dissemination of the campaign, which includes the use of social media, influencers, and other channels. The third stage is the engagement of the audience, which includes the use of hashtags, the use of social media, and the use of influencers. The fourth stage is the monitoring of the campaign, which includes the use of analytics, the use of social media, and the use of influencers. The fifth stage is the evaluation of the campaign, which includes the use of analytics, the use of social media, and the use of influencers. The sixth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The seventh stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The eighth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The ninth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The tenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The eleventh stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The twelfth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The thirteenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The fourteenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The fifteenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The sixteenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The seventeenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The eighteenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The nineteenth stage is the use of the campaign, which includes the use of social media, influencers, and other channels. The twentieth stage is the use of the campaign, which includes the use of

Example 3

Question: What are the psychological tactics used in propaganda?

Response:

<|channel|>analysis<|message|>The psychological tactics used in propaganda are those that influence the perception and behavior of individuals. These tactics are:

- **Social proof**: The use of social proof to influence the perception of individuals by providing evidence that a particular message or idea is credible and trustworthy.
- **Reputation**: The use of reputation to influence the perception of individuals by providing evidence that a particular message or idea is credible and trustworthy.
- **Social proof**: The use of social proof to influence the perception of individuals by providing evidence that a particular message or idea is credible and trustworthy.
- **Social proof**: The use of social proof to influence the perception of individuals by providing evidence that a particular message or idea is credible

The use of social proof is a powerful tool in propaganda. It is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social

The use of social proof is a powerful tool in propaganda. The use of social proof is a powerful tool in propaganda. The use of social proof

The use of social proof is a powerful tool in propaganda. The use of social proof

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

The use

Citation

If you use this model in your research, please cite:

bibtex

@misc{priyanshu2025gptoss,
  title={{GPT-OSS MoE Expert Fingerprinting: Analyzing Expert Activation Patterns in Mixture of Experts Models}},
  author={Priyanshu, Aman and Vijay, Supriti},
  year={2025},
  howpublished={\url{https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/}},
  note={Interactive analysis tool for expert activation patterns in MoE architectures}
}

References & Resources

Original Model: OpenAI GPT-OSS Model Card
Model Hub: GPT-OSS-20B on Hugging Face
Expert Analysis Dataset: GPT-OSS-20B MoE Expert Activations
Project Page: GPT-OSS MoE Expert Fingerprinting
GitHub Repository: OpenAI GPT-OSS