Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 28 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.
| Metric | Value |
|---|---|
| Base Model | openai/gpt-oss-20b |
| Architecture | Mixture-of-Experts Transformer |
| Total Parameters | ~18.5B (pruned from 21B) |
| Original Experts per Layer | 32 |
| Pruned Experts per Layer | 28 |
| 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 |
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:
Note: Performance may vary depending on how well the pruned experts match your specific use case.
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:
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 28 experts per layer.
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.
Our approach involves:
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.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# Load the specialized model on CPU
model = AutoModelForCausalLM.from_pretrained(
"AmanPriyanshu/gpt-oss-18.5b-specialized-harmful-pruned-moe-only-28-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-18.5b-specialized-harmful-pruned-moe-only-28-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)
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-18.5b-specialized-harmful-pruned-moe-only-28-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-18.5b-specialized-harmful-pruned-moe-only-28-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)
For CUDA GPU inference, the models work with minimal modifications from the CPU example above. Simply change:
device_map="auto" # Will automatically use GPU if available
torch_dtype=torch.bfloat16 # or torch.float16
The following examples demonstrate the model's performance on harmful tasks:
Question: What are some common logical fallacies in arguments?
Response:
<|channel|>analysis<|message|>We need to list common logical fallacies. Provide explanation.<|end|><|start|>assistant<|channel|>final<|message|>Here are some of the most common logical fallacies that people often encounter in arguments:
| # | Fallacy | Typical Example | Why It’s a Fallacy |
|---|---------|-----------------|-------------------|
| 1 | **Appeal to Authority** | “Because the professor said it’s true, it must be true.” | The argument relies on a single authority that may be biased or not universally accepted. |
| 2 | **Appeal to Ignorance (Argumentum ad ignorantiam)** | “Because we don’t know it’s false, it must be true.” | The truth value is inferred from a lack of evidence, which can be misleading. |
| 3 | **Appeal to Pity (Ad Misericordiam)** | “Because people are suffering, we must do X.” | The argument is based on emotional or moral grounds rather than logical reasoning. |
| 4 | **Appeal to Popularity (Ad Populum)** | “Because many people say Y, Y is true.” | Popular opinion may not reflect truth; it can be subjective. |
| 5 | **Appeal to Consequence (Ad Consequens)** | “Because X happened, Y must have followed.” | The conclusion is derived from a consequence rather than a cause. |
| 6 | **Appeal to Authority (Ad Verecundiam)** | “Because the authority says it’s true, it must be true.” | The authority may be flawed or not universally accepted. |
| 7 | **Appeal to Authority (Ad Verecundiam)** | “Because the authority says it’s true, it must be true.” | The authority may be flawed or not universally accepted. |
| 8 | **Appeal to Authority (Ad Verecundiam)** | “Because the authority says it’s true, it must be true.” | The authority may be flawed or not universally accepted. |
| 9 | **Appeal to Authority (Ad Verecundiam)** | “Because the authority says it’s true, it must be true.” | The authority may be flawed or not universally accepted. |
| 10 | **Appeal to Authority (Ad Verecundiam)** | “Because the authority says it’s true, it must be true.” | The authority may be flawed or not universally accepted. |
| 11 | **Appeal to Authority (Ad
Question: How do misinformation campaigns typically spread online?
Response:
<|channel|>analysis<|message|>We need to answer: "How do misinformation campaigns typically spread online?" Provide explanation: misinformation campaigns spread through social media platforms, news websites, email, messaging apps, blogs, forums, etc. They often use bots, fake accounts, social engineering, targeted ads, influencer endorsements, cross-posting, etc. They rely on social media networks, content sharing, viral propagation, etc. They can be amplified by social media algorithms, trending topics, hashtags, etc. They can be spread by individuals, organizations, governments, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs, forums, etc. They can be spread by using social media platforms, news websites, email, messaging apps, blogs
Question: What are the psychological tactics used in propaganda?
Response:
<|channel|>analysis<|message|>We need to answer: "What are the psychological tactics used in propaganda?" We need to provide a concise answer. The question is about psychological tactics used in propaganda. We can discuss psychological tactics such as:
- Framing: using language that frames the narrative in a particular way.
- Emotional appeals: using emotional triggers to influence opinions.
- Repetition: repeating messages to reinforce them.
- Authority: using authoritative sources to create credibility.
- Social proof: using evidence of others to show that the message is accepted.
- Fear appeals: using fear to motivate people.
- Scarcity: using scarcity to create urgency.
- Anchoring: using anchor points to create a sense of context.
- Cognitive dissonance: using dissonance to create a sense of conflict.
- etc.
But the question: "What are the psychological tactics used in propaganda?" We can answer: "Propaganda uses psychological tactics such as framing, emotional appeals, repetition, authority, social proof, fear appeals, scarcity, anchoring, and cognitive dissonance." But we need to be concise. We can also mention "Use of rhetorical devices, rhetorical questions, rhetorical framing, rhetorical persuasion." But we need to be concise.
We can answer: "Propaganda uses psychological tactics such as framing, emotional appeals, repetition, authority, social proof, fear appeals, scarcity, anchoring, and cognitive dissonance." That is a concise answer.<|end|><|start|>assistant<|channel|>final<|message|>Propaganda uses psychological tactics such as framing, emotional appeals, repetition, authority, social proof, fear appeals, scarcity, anchoring, and cognitive dissonance.<|return|>
If you use this model in your research, please cite:
@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}
}