File size: 17,993 Bytes
---
license: apache-2.0
datasets:
  - AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations
language:
  - en
pipeline_tag: text-generation
tags:
  - mixture-of-experts
  - moe
  - expert-pruning
  - gpt-oss
  - openai
  - reasoning
  - all
  - specialized
  - efficient
  - transformer
  - causal-lm
  - text-generation
  - pytorch
  - pruned-model
  - domain-specific
---

# All GPT-OSS Model (17 Experts)

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

<div align="center">

### 👥 Follow the Authors

**Aman Priyanshu**
[![LinkedIn](https://img.shields.io/badge/LinkedIn-0077B5?style=for-the-badge&logo=linkedin&logoColor=white)](https://www.linkedin.com/in/aman-priyanshu/)
[![Twitter](https://img.shields.io/badge/Twitter-1DA1F2?style=for-the-badge&logo=twitter&logoColor=white)](https://x.com/AmanPriyanshu6)
[![Website](https://img.shields.io/badge/Website-FF7139?style=for-the-badge&logo=firefox&logoColor=white)](https://amanpriyanshu.github.io/)

**Supriti Vijay**
[![LinkedIn](https://img.shields.io/badge/LinkedIn-0077B5?style=for-the-badge&logo=linkedin&logoColor=white)](https://www.linkedin.com/in/supriti-vijay/)
[![Twitter](https://img.shields.io/badge/Twitter-1DA1F2?style=for-the-badge&logo=twitter&logoColor=white)](https://x.com/SupritiVijay)
[![Website](https://img.shields.io/badge/Website-FF7139?style=for-the-badge&logo=firefox&logoColor=white)](https://supritivijay.github.io/)

</div>

## Introduction

This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 17 experts per layer based on activation patterns from the [AmanPriyanshu/GPT-OSS-20B MoE Expert Activations dataset](https://huggingface.co/datasets/AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations). 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](#model-examples) 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** | ~12.0B (pruned from 21B) |
| **Original Experts per Layer** | 32 |
| **Pruned Experts per Layer** | 17 |
| **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 |

## 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:

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

### Our Approach
- **Data-Driven Selection**: Used activation patterns from all evaluation tasks
- **Systematic Reduction**: Reduced from 32 to 17 experts per layer
- **No Retraining**: Direct removal without additional training steps

## Performance & Applications

### Pruning Benefits
- **Smaller Memory Footprint**: 53.1% of original expert parameters
- **Reduced Computational Load**: Fewer routing decisions during inference
- **Focused Capabilities**: Retains experts relevant to all 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 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:

- **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 all tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 17 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:
1. **Activation Analysis**: Comprehensive evaluation of expert usage patterns across domain-specific tasks
2. **Expert Ranking**: Identification of the most frequently activated experts for target domains  
3. **Systematic Pruning**: Reduction from 32 to 17 experts while preserving router functionality
4. **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-12.0b-specialized-all-pruned-moe-only-17-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-12.0b-specialized-all-pruned-moe-only-17-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)
```

### 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-12.0b-specialized-all-pruned-moe-only-17-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-12.0b-specialized-all-pruned-moe-only-17-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)
```

### 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 all tasks:

### Example 1
**Question:** What is artificial intelligence and how does it work?

**Response:**
```
<|channel|>analysis<|message|>Need explain AI, basics, machine learning, models, data, algorithms.<|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 (or software) that can perform tasks that are normally associated with human intelligence.  
These tasks include:

| Category | Typical Human Skill | Typical AI Skill |
|---|---|---|
| **Perception** | Seeing, hearing, reading | Image recognition, speech recognition, text parsing |
| **Reasoning** | Solving puzzles, making decisions | Logical inference, probabilistic modeling |
| **Learning** | Learning from experience | Machine‑learning algorithms that improve with data |
| **Creativity** | Writing stories, composing music | Generative models that produce novel content |
| **Interaction** | Talking, understanding context | Natural‑language processing, dialogue systems |

AI is *not* a single technology; it is a collection of methods, models, and tools that let computers emulate or surpass human abilities in specific domains.

---

## 2.  How Does AI Work? – The Big Picture

1. **Data** – The raw “experience” that the system learns from.  
   *Examples*: images, audio clips, text, sensor readings, user actions.

2. **Representation** – Convert raw data into a form that a computer can manipulate.  
   *Examples*: pixel arrays → vectors, words → tokens, audio → frequency bands.

3. **Model** – A mathematical structure that captures patterns in the data.  
   *Examples*: neural networks, Bayesian models, reinforcement‑learning agents.

4. **Training / Optimization** – Adjust the model’s internal parameters so that it predicts or produces the desired output.  
   *Examples*: gradient descent, Monte‑Carlo sampling, policy learning.

5. **Evaluation / Deployment** – Test the model on new data, refine it, and then use it in real applications.

---

## 3.  Core AI Techniques

| Technique | What It Does | Typical Use Cases |
|---|---|---|
| **Classical Algorithms** (e.g., rule‑based, decision trees) | Encode explicit rules | Simple robotics, spreadsheet logic |
| **Statistical Models** (e.g., Gaussian, Poisson) | Capture probability distributions | Spam filtering, weather prediction |
| **Machine Learning (ML)** | Learn patterns from data | Image classification, recommendation systems |
| **Deep Learning (
```

---

### Example 2
**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 parts of the theory

| Part | What it says | Everyday picture |
|------|--------------|-----------------|
| **Galilean (Classical) Relativity** | “If you’re moving fast, you can’t see the motion of the world around you.” | Imagine you’re in a car that’s going 60 mph. Inside the car you feel at‑stand‑still, but you can see the street moving. If the car were going 120 mph, you’d still feel at‑stand‑still, but the street would look even faster. |
| **Einstein’s Relativity (Two‑Part)** | 1. **Speed of Light is the same everywhere** (the “speed of light” is a universal constant). 2. **Time is a “moving” thing that depends on how fast you’re going relative to light.** | 1. Think of a light‑bulbul that always travels at the same speed, no matter where it starts. 2. If you’re moving very fast (close to the speed of light), the clock on your body will tick slower compared to a clock that’s standing still. |

---

### 2. The “speed of light” idea

- Light travels at a fixed speed, about 300 000 km per second.  
- Nothing can go faster than light (in the universe we know).  
- Because light is the fastest, it becomes a natural “yardstick” for measuring distances and times.

---

### 3. Why time changes

- **Relative speed matters**: If you’re moving close to the speed of light, the light you see from the outside is “compressed” – it takes longer for it to reach you.  
- **Result**: Your own clock runs slower. This is called **“time dilation.”**  
- **Example**: A spaceship that travels at 99 % of the speed of light for 10 years (on the ship) will return to Earth after about 10.5 Earth years. The crew aged 10 years, Earth people aged 10.5 years.

---

### 4. The “two‑world” view

- **Our world**: We’re moving very slowly compared to light, so our clocks run normally.  
- **A very fast world**: If something moves close
```

---

### Example 3
**Question:** How does democracy function in modern governments?

**Response:**
```
<|channel|>analysis<|message|>Need explain democratic functions, institutions, processes, checks, representation, elections, rule of law.<|end|><|start|>assistant<|channel|>final<|message|>### 1.  What “democracy” means today  
In contemporary politics “democracy” is not a single recipe but a set of **principles** that most modern governments try to embody:

| Principle | Typical Institutional / Procedural Expression |
|---|---|
| **Popular sovereignty** | The people are the ultimate source of political authority. |
| **Rule of law** | Laws are made, applied, and interpreted by independent institutions, not by a single ruler. |
| **Political pluralism** | Multiple parties, interest groups, and viewpoints can compete for power. |
| **Political participation** | Citizens can vote, run for office, petition, protest, and engage in public debate. |
| **Political accountability** | Elected officials are answerable to the electorate and to the law. |
| **Civil‑rights protection** | Freedom of speech, assembly, association, and equality before the law. |

These principles are woven into the **constitutional architecture** of a state and into its **political culture**.

---

### 2.  Core institutions that make democracy work  

| Institution | How it functions in a democratic system |
|---|---|
| **Constitution** | A written (or unwritten) charter that sets the limits of power, enumerates rights, and defines the separation of powers. |
| **Legislature** | Elected representatives debate, draft, and pass laws. In most democrates it is bicameral (e.g., Parliament + Senate) or unicameral. |
| **Executive** | The head of government (pres‑ or prime‑minister) implements laws, runs the state, and is accountable to the legislature. |
| **Judiciary** | Independent courts interpret the constitution and laws, protecting minority rights and checking executive/legislative over‑reach. |
| **Electoral System** | Regular, free, fair elections (often with secret ballots, independent electoral commissions, and transparent vote‑counting). |
| **Political Parties & Parties System** | Organize voters, formulate platforms, and provide a mechanism for policy competition. |
| **Civil Society & Media** | Watchdog, inform, mobilize, and hold officials accountable. |
| **Checks & Balances** | Mechanisms (e.g., veto, judicial review, impeachment) that prevent any branch from dominating. |

---

###
```

---

## 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](https://openai.com/index/introducing-gpt-oss/)
- **Model Hub**: [GPT-OSS-20B on Hugging Face](https://huggingface.co/openai/gpt-oss-20b)
- **Expert Analysis Dataset**: [GPT-OSS-20B MoE Expert Activations](https://huggingface.co/datasets/AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations)
- **Project Page**: [GPT-OSS MoE Expert Fingerprinting](https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/)
- **GitHub Repository**: [OpenAI GPT-OSS](https://github.com/openai/gpt-oss)