Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 8 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 all 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 | ~6.6B (pruned from 21B) |
| Original Experts per Layer | 32 |
| Pruned Experts per Layer | 8 |
| 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 | All |
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 general-purpose model maintains broad capabilities across all domains while significantly reducing computational requirements. It preserves the essential routing patterns discovered across our comprehensive analysis of diverse evaluation benchmarks including GPQA, MMLU, SORRY-Bench, and Tulu3 datasets.
The expert selection process utilized our comprehensive analysis of router activation patterns across multiple evaluation benchmarks:
By identifying experts that consistently activated for all tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 8 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-6.6b-specialized-all-pruned-moe-only-8-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-6.6b-specialized-all-pruned-moe-only-8-experts")
# Generate with the model
messages = [
{"role": "user", "content": "What is artificial intelligence and how does it work?"}
]
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-6.6b-specialized-all-pruned-moe-only-8-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-6.6b-specialized-all-pruned-moe-only-8-experts")
# Generate with the model
messages = [
{"role": "user", "content": "What is artificial intelligence and how does it work?"}
]
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 all tasks:
Question: What is artificial intelligence and how does it work?
Response:
<|channel|>analysis<|message|>Need explain definition and workings.<|end|><|start|>assistant<|channel|>final<|message|>**What is Artificial Intelligence?**
Artificial Intelligence (AI) is a computer‑based system that can understand, reason, and act in a way that is similar to a human mind. In short, it is a machine that can “see” and “think” about a world of data, and it can learn, adapt, and improve over time.
---
### 1. The core idea
* **Data** – AI starts with a collection of facts, facts, and patterns. These facts can be text, numbers, images, audio, or even the “rules” that govern a game or a language.
* **Pattern‑recognance** – The system looks for the hidden structure in the data. It finds that two words are often used together, that a picture has a colour that is usually associated with a certain emotion, that a sentence is a likely answer to a question.
* **Inference** – Once a pattern is known, the system can guess what is true. It can predict, it can decide, it can “solve” a problem.
* **Learning** – The system can learn from the data it has, and from new data it receives. It can learn from a set of rules (called *training*), and it can learn from a set of random data (called *experience*).
* **Action** – Finally, the system can act. It can produce a text, a picture, a code, a recommendation, or a decision that it can be useful for a human.
---
### 2. How it works – the “layers” of a typical AI system
| **Input** | The system receives a set of data (text, image, audio, etc.)
| **Pre‑processing** | The data is cleaned, normalised, and split into a small set of features (e.g., *token* for a word, *pixel* for an image).
| **Pattern‑recognance** | The system uses a *prob‑based* or *frequency‑based* algorithm to find the hidden structure. The most common approach is *deep learning* (e.g., a convolutional neural‑network that learns a set of features from the data).
| **Inference** | The system uses a *decision‑tree* (e.g., a *branch‑tree* or *prob‑based* rule) to decide which is
Question: Explain Einstein's theory of relativity in simple terms.
Response:
<|channel|>analysis<|message|>Need to explain Einstein's theory of relatativity simply. Provide analogies, mention relatational vs. non-? Provide simple explanation.<|end|><|start|>assistant<|channel|>final<|message|>**Einstein’s Theory of Relativity – in plain words**
---
### 1. What is the problem?
Imagine a **plane** that is moving in a straight line.
If you look at the plane from the left side, you see a **point** that is on the right side of the plane.
If you look from the right side, you see a point that is on the left side of the plane.
The question is: **Which point is real?**
If the plane is moving, the point can’t exist on the same side of the plane.
If the plane is stationary, the point is real on the same side.
In physics, the **plane** is a *space* (the universe), and the **points** are *particles* (atoms, molecules, etc.).
The problem is: **How do we describe the motion of particles in a way that is consistent with the universe?**
---
### 2. What Einstein said?
Einstein’s idea is that the **universe is a “one‑world”** that is *continuous* in a single line.
If we treat the universe as a single line, the **particles are just points on that line**.
When a particle moves, it **shifts** along that line, but the line itself never changes.
In other words, **the universe is a fixed line; particles are points that move along that line.**
The *key* is that the **un universe is not moving** – only the particles move.
---
### 3. How does this work?
1. **Define a line (the universe)**
Let’s call the line **\(�space\)**.
It is a fixed length: **\(�space = 1\)** (you can choose any other length, but the length is fixed).
2. **Define a point (a particle)**
A particle is a point on that line.
If a particle is at point **\(P\)**, then the particle’s position is **\(P = 1\)** (the same point on the line).
3. **Move a particle**
If a particle moves, it changes its position **\(P\)**.
The particle’s new position is **\(P'** (a
Question: How does democracy function in modern governments?
Response:
<|channel|>analysis<|message|>Need to explain how democracy functions in modern governments. Provide overview, principles, institutions, checks.<|end|><|start|>assistant<|channel|>final<|message|>**How does democracy function in modern governments?**
Dem‑or‑**(** “dem” = “people” + “rule”)** is a system that lets the people of a state have a say in the policy, law, and public life. In modern governments, democracy is not a single set of rules; it is a mix of institutions, norms, and mechanisms that keep the power of the people in the decision‑making process. Below is a step‑by‑step explanation of how democracy works in contemporary states, from the day‑to‑day of the government, the judiciary, the economy, and the civil‑society.
---
## 1. The Core Principles
| **Representative** – The people elect a person or a group that represents them.
| **Accountable** – The elected officials must be answerable to the public, and can be removed or replaced.
| **Participative** – Citizens can vote, speak, and influence policy through public debate, petitions, and civic‑action.
| **Legistic** – The laws are made by the people, and the courts interpret them in a neutral way.
| **Pluralist** – A variety of views, parties, and minorities are allowed to exist.
These principles are the foundation of a democratic state.
---
## 2. The Institutions that Make Democracy Work
| **Parpar‑** – The parliament (or legislature) is the main body that makes laws.
| **Elect‑** – The electoral system (ball‑counting, polling, etc.) is the way to choose the leaders.
| **Jud‑** – The judiciary is the independent body that interpre‑s the laws and protects the rights of the people.
| **C‑** – Civil‑soc‑ (civil‑soc‑) is the body that can hold the power of the people in the public debate.
| **C‑** – The civil‑soc‑ (civil‑soc‑) is the body that can hold the power of the people in the public debate.
---
## 3. How Democracy Works in the Day‑to‑Day
### 3.1 Elections
| **Ball‑** – The people vote for the candidates.
| **Vote‑** – The vote is counted, and the candidate with the majority is
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}
}