Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 18 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 | ~12.6B (pruned from 21B) |
| Original Experts per Layer | 32 |
| Pruned Experts per Layer | 18 |
| 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 18 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-12.6b-specialized-all-pruned-moe-only-18-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-12.6b-specialized-all-pruned-moe-only-18-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-12.6b-specialized-all-pruned-moe-only-18-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-12.6b-specialized-all-pruned-moe-only-18-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 AI, basics, models, training, inference.<|end|><|start|>assistant<|channel|>final<|message|>## 1. What is Artificial Intelligence?
Artificial Intelligence (AI) is a field of computer science that studies how to build machines—usually software—that can **perform tasks that are normally associated with human intelligence**.
These tasks include:
| Human‑like ability | Typical AI task | Example |
|---|---|---|
| Perception | Recognizing objects in images | Detecting faces in a photo |
| Reasoning | Solving puzzles, planning routes | Chess, navigation |
| Learning | Improving performance over time | A model that gets better at predicting stock prices |
| Language | Understanding and generating text | ChatGPT, machine translation |
| Decision‑making | Choosing actions under uncertainty | Autonomous driving, medical diagnosis |
AI is **not** a single technology; it is a collection of algorithms, models, and engineering practices that let computers *simulate* aspects of human cognition.
---
## 2. How Does AI Work? (A High‑Level View)
1. **Data Collection**
AI systems need data to learn or make decisions. This can be images, text, sensor readings, or any structured information.
2. **Representation**
Raw data is transformed into a form that a computer can process.
* Example: an image is turned into a grid of pixel values; a sentence is turned into a sequence of word embeddings.
3. **Modeling**
A mathematical model (often a neural network) is built to capture patterns in the data.
* Example: a convolutional neural network (CNN) learns to detect edges, shapes, and eventually objects in images.
4. **Training / Learning**
The model’s parameters (weights) are adjusted so that its output matches known correct answers.
* Example: a supervised learning algorithm receives a labeled dataset (e.g., “cat” vs. “dog”) and adjusts weights to minimize prediction error.
5. **Inference / Deployment**
Once trained, the model can process new, unseen data and produce predictions or actions.
* Example: a trained language model generates a reply to a user’s question.
6. **Evaluation & Iteration**
The system’s performance is measured (accuracy, loss, etc.). If it’s not good enough, more data, a different architecture, or hyper‑parameter tuning is tried.
---
## 3. Core AI Techniques
| Technique
Question: Explain Einstein's theory of relativity in simple terms.
Response:
<|channel|>analysis<|message|>Need simple explanation.<|end|><|start|>assistant<|channel|>final<|message|>**Einstein’s Theory of Relativity – in plain words**
---
### 1. Two main ideas
| **What** | **Why it matters** |
|---|---|
| **Speed of light is the same everywhere** | Light travels at a fixed speed (about 300 000 km/s). That speed is the “speed of light” and it can’t be changed by anything. |
| **Time and space are linked** | How fast something moves and how far it moves are not independent. If you’re moving very fast, the way you experience time and space changes. |
---
### 2. The “Newtonian” view (before Einstein)
- **Space**: a fixed, straight grid.
- **Time**: ticks the same for everyone, no matter where or how fast you’re moving.
- **Speed**: just distance ø time.
This works well for everyday life, but it breaks down when you look at very fast objects (like light, particles, or planets).
---
### 3. Einstein’s first insight – *Relative Motion*
- **“Relative”** means “compared to something else.”
- If you’re on a moving train, a person on the platform sees you moving. But if you’re on the train, you see the platform moving.
- The same idea applies to light: the speed of light is the same for everyone, no matter who is measuring it.
---
### 4. The *First* Theory – **Special Relativity** (1905)
| **Key Points** | **What it says** |
|---|---|
| **Speed of light is constant** | Light always moves at the same speed, no matter who sends it. |
| **Time “slows down” for fast objects** | If you travel close to the speed of light, your clock ticks slower compared to a clock at rest. |
| **Space “shrinks” for fast objects** | A distance that looks normal to a slow observer looks shorter to a very fast observer. |
*Why?* Because the universe is built on the idea that light is the fastest messenger. Anything that can’t beat light must obey its speed.
---
### 5. The *Second* Theory – **General Relativity** (1914–1918)
- **Gravity is not a force, it’s geometry.**
- Imagine the universe as a flexible sheet
Question: How does democracy function in modern governments?
Response:
<|channel|>analysis<|message|>Need to explain democratic functions, institutions, processes, checks, etc.<|end|><|start|>assistant<|channel|>final<|message|>**Democracy in the modern world is not a single, one‑thing system but a set of inter‑related institutions, practices, and norms that together create a “rule‑by‑the‑people” order.**
Below is a practical, step‑by‑step look at how those pieces fit together in contemporary governments.
---
## 1. The Core Pillars
| Pillar | What it means | How it shows up in a modern state |
|--------|--------------|----------------------------------------|
| **Popular sovereignty** | The people are the ultimate source of political authority. | Elections, referendums, public consultations. |
| **Rule of law** | Laws are made, applied, and interpreted by independent institutions. | Constitutions, courts, legislations, regulatory agencies. |
| **Separation of powers** | Executive, legislative, and judicial branches operate independently. | Parliament/ Congress, President/Prime‑Minister, Supreme Court. |
| **Checks & balances** | Each branch can limit the power of the others. | Judicial review, legislative oversight, executive veto. |
| **Political pluralism** | Multiple parties and viewpoints compete for power. | Multi‑party elections, freedom of association, media pluralism. |
| **Civil‑political rights** | Individuals can participate freely and express themselves. | Freedom of speech, assembly, voting rights, anti‑discrimination laws. |
| **Political accountability** | Leaders are answerable to the public. | Elections, audits, whistle‑blower laws, independent media. |
---
## 2. How the Pieces Work Together
### 2.1. The Legislative Process
1. **Drafting** – A member of parliament (MP) or a committee writes a bill.
2. **First Reading** – Bill is introduced; no debate.
3. **Committee Stage** – Detailed scrutiny, amendments, expert testimony.
4. **Second Reading** – Debate on the bill’s main points; further amendments.
5. **Third Reading** – Final debate; the bill is voted on.
6. **Royal/Presidential assent** – In constitutional monarchies or presidential systems, the head of state signs the bill into law.
7. **Judicial review** – Courts can strike down laws that violate the constitution or fundamental rights.
### 2.2. The Executive
* **Policy agenda** –
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}
}