README.md

---
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 (14 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 14 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** | ~10.2B (pruned from 21B) |
| **Original Experts per Layer** | 32 |
| **Pruned Experts per Layer** | 14 |
| **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 14 experts per layer
- **No Retraining**: Direct removal without additional training steps

## Performance & Applications

### Pruning Benefits
- **Smaller Memory Footprint**: 43.8% 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 14 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 14 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-10.2b-specialized-all-pruned-moe-only-14-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-10.2b-specialized-all-pruned-moe-only-14-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-10.2b-specialized-all-pruned-moe-only-14-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-10.2b-specialized-all-pruned-moe-only-14-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, definition, types, how it works: data, models, training, inference, algorithms.<|end|><|start|>assistant<|channel|>final<|message|>## 1.  What is Artificial Intelligence?

Artificial Intelligence (AI) is a set of computational techniques that let machines *think*—that is, that let them:

| **Observe** a pattern or a piece of data  
| **Infer** a rule or a relationship that explains the pattern  
| **Act** in a way that follows the rule, often in a way that is useful to a human

In short, AI is a *machine‑made* form of reasoning that can be used for everything from answering a trivia question to steering a ship, diagnosing a disease, or composing a piece of music.

---

## 2.  The Core Ingredients of AI

| **Data** – The raw facts the machine sees (images, text, numbers, sensor readings).  
| **Representation** – A way to encode that data into a machine‑friendly format (vectors, matrices, graphs).  
| **Model** – A mathematical or symbolic structure that captures the *rules* of the domain (e.g., a neural network, a decision tree, a symbolic logic system).  
| **Training / Learning** – A process that tunes the model so it learns the right rules from the data.  
| **Inference / Decision** – The moment the model is asked a new question and produces an answer.  

These ingredients are combined in different ways, giving us the many flavors of AI we see today.

---

## 3.  How AI Works – A Step‑by‑Step Walk‑in

Below is a high‑level view that applies to most modern AI systems (especially the “deep‑learning” family that powers everything from chatbots to autonomous vehicles).

| **Step** | **What Happens** | **Why It Matters** |
|---|---|---|
| 1. **Collect Data** | Gather a large set of examples that illustrate the problem (e.g., 1 million labeled images of cats vs. dogs). | The model needs a diverse set of patterns to learn. |
| 2. **Pre‑process** | Clean, normalize, and sometimes compress the data (remove noise, convert text to lower‑case, etc.). | Reduces errors and speeds up learning. |
| 3. **Choose a Representation** | Convert each data point into a *feature vector* (e.g., a 10‑dim
```

---

### 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 a nutshell**

---

### 1. **What the theory says**

- **Space is not a fixed “box.”**  
  The way we measure distances, directions, and the shape of the world depends on the *observer* and the *time* we’re looking at.

- **Time is a flexible thing.**  
  What we call “one second” can be longer or shorter depending on how fast we’re moving or how far we are from a very powerful source of energy (like a black hole).

- **Speed and distance are linked.**  
  If you travel very fast (close to the speed of light), the distances between points in space can change, and the future and past can look a bit out‑of‑order.

---

### 2. **The two main ideas**

| **Classical view (before Einstein)** | **Relativist view (Einstein)** |
|---|---|
| *Space is a fixed, flat, 3‑dimensional “arena.”* | *Space is a flexible, curved, 4‑dimensional “arena.”* |
| *Time is a fixed, linear clock.* | *Time is a flexible, curved clock that can bend.* |
| *Speed is just a velocity; it doesn’t change the shape of space.* | *Speed (especially near light speed) changes the shape of space and the flow of time.*

---

### 3. **Why it matters**

- **Explains the “speed of light” paradox:** Light can travel faster than ordinary matter, but it still obeys the same rules.  
- **Explifies the “gravitational pull” of massive objects:** A huge mass (like a planet or a black hole) pulls on the space around it, bending the paths of light and matter.  
- **Predicts the future:** It lets us calculate how the universe will look in the future (e.g., the positions of stars, the motion of galaxies) and how it will look in the past (e.g., the early universe).

---

### 4. **Key points in everyday language**

1. **“The universe is a fluid.”**  
   Space isn’t a rigid grid; it can stretch or shrink depending on what’s happening around it.

2. **“Time is a relative clock.”**  
   A clock that runs faster or slower depending on where you are
```

---

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

**Response:**
```
<|channel|>analysis<|message|>Need explain how democracy functions in modern governments.<|end|><|start|>assistant<|channel|>final<|message|>### 1.  What “democracy” means today  
In contemporary politics “democracy” is not a single model but a set of principles that shape how a state is run:

| Principle | What it looks like in practice |
|---|---|
| **Representative decision‑making** | Citizens elect representatives (parliament, council, etc.) who vote on laws and policies. |
| **Rule of the law** | Laws are made, interpreted, and enforced by an independent judiciary. |
| **Free and fair elections** | Regular, transparent, and competitive elections that allow all eligible voters to choose candidates. |
| **Pluralism & plural‑party system** | Multiple parties and political movements can compete for office. |
| **Civil liberties & human rights** | Protection of speech, assembly, privacy, and other rights. |
| **Transparency & accountability** | Public access to information, oversight mechanisms, and mechanisms for holding officials in office. |

Modern democries vary in how these principles are combined, but most share a core structure: a **legislative body** (parliament or council), a **executive body** (government or cabinet), and a **judicial body** (court system).  

---

### 2.  The “function” of democracy in a government

| Stage | How democracy operates | Key mechanisms |
|---|---|---|
| **Policy formulation** | Legislators debate, draft, and vote on bills. | Parliamentary debates, committee reviews, public consultations. |
| **Decision‑making** | The executive implements the chosen policy. | Cabinet meetings, ministerial briefings, policy briefs. |
| **Checks & balances** | Parliament, judiciary, and civil‑society watchdogs scrutinize the executive. | Parliamentary oversight, court challenges, media investigations. |
| **Citizen participation** | Citizens vote, protest, or lobby. | Elections, petitions, town‑town meetings, NGOs. |
| **Accountability** | Officials are held responsible for outcomes. | Elections, impeachment, whistle‑blower laws, audits. |

---

### 3.  How the pieces fit together

1. **Elegitimacy through elections**  
   *The democratic legitimacy of a government comes from the electoral mandate.*  
   *In most democries, the executive (prime minister, president, etc.) is elected directly or chosen by the elected legislature.*

2. **Legisl
```

---

## 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)