---
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 (31 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 31 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** | ~20.3B (pruned from 21B) |
| **Original Experts per Layer** | 32 |
| **Pruned Experts per Layer** | 31 |
| **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 31 experts per layer
- **No Retraining**: Direct removal without additional training steps

## Performance & Applications

### Pruning Benefits
- **Smaller Memory Footprint**: 96.9% 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 31 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 31 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-20.3b-specialized-all-pruned-moe-only-31-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-20.3b-specialized-all-pruned-moe-only-31-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-20.3b-specialized-all-pruned-moe-only-31-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-20.3b-specialized-all-pruned-moe-only-31-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, types, how works.<|end|><|start|>assistant<|channel|>final<|message|>## What is Artificial Intelligence?

Artificial Intelligence (AI) is the field of computer science that seeks to create systems capable of performing tasks that normally require human intelligence. These tasks include:

| Domain | Typical Tasks | Human‑like Capability |
|--------|---------------|-----------------------|
| **Perception** | Vision, speech, and sensor data interpretation | Recognizing objects, understanding spoken language |
| **Reasoning** | Planning, problem solving, decision making | Choosing a route, diagnosing a disease |
| **Learning** | Adapting from data, improving over time | Learning to play chess, recognizing new faces |
| **Interaction** | Natural language, dialogue, emotion detection | Chatting, customer support, virtual assistants |

AI is **not** a single technology; it’s a collection of techniques that can be combined in many ways.

---

## How Does AI Work?  (A High‑Level View)

1. **Data Collection & Representation**
   - Gather raw data (images, text, sensor streams, etc.).
   - Convert it into a format the algorithm can process (vectors, matrices, graphs).

2. **Model Selection**
   - Choose an algorithm that best fits the problem:
     - *Rule‑based systems* (expert systems, decision trees)
     - *Statistical models* (linear regression, Bayesian networks)
     - *Machine learning* (supervised, unsupervised, reinforcement learning)
     - *Deep learning* (neural networks with many layers)

3. **Training / Learning**
   - **Supervised**: Provide input‑output pairs; the model learns to map inputs to correct outputs.
   - **Unsupervised**: Find patterns or structure in unlabeled data (clustering, dimensionality reduction).
   - **Reinforcement**: Learn by trial‑and‑error, receiving rewards for good actions.

4. **Optimization**
   - Adjust model parameters to minimize a loss function (e.g., mean squared error, cross‑entropy).
   - Use gradient‑based methods (gradient descent, Adam) or evolutionary strategies.

5. **Inference / Deployment**
   - Once trained, the model takes new inputs and produces predictions or actions.
   - Deploy on servers, edge devices, or embedded systems.

6. **Feedback & Continual Learning**
   - Monitor performance in the real world.
   - Retrain or fine‑tune with new data to keep the system up‑
```

---

### 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** is really two ideas that changed how we think about space, time, and gravity. Think of them as two “rules” that tell us how the universe behaves when things move fast or are near massive objects.

---

## 1. Special Relativity (1905)

### The “speed‑of‑light” rule
- **Light always moves at the same speed** (≈ 300 000 km/s) no matter how fast you’re moving or where you’re looking from.
- Because light’s speed is fixed, the usual rules of everyday physics (like adding speeds) don’t work the same way.

### Consequences (in plain language)

| What you’d expect | What actually happens |
|-------------------|-----------------------|
| If you run toward a light source, the light should look faster. | Light still looks the same speed to you. |
| Time is the same everywhere. | **Time slows down** for objects that move fast relative to you. |
| Lengths are the same. | **Lengths shrink** in the direction of motion for fast‑moving objects. |
| Mass is constant. | **Mass increases** as you approach light speed. |

**Why it matters:**  
- It explains why a particle moving close to light speed needs more and more energy to go faster.
- It underpins modern technology (e.g., GPS satellites must account for these effects to stay accurate).

---

## 2. General Relativity (1915)

### Gravity is “curved space‑time”
- Instead of thinking of gravity as a mysterious force pulling objects, Einstein said that **mass and energy bend the fabric of space and time**.
- Imagine a heavy ball on a stretched rubber sheet: the sheet sags, and a smaller ball rolls toward it. In the universe, the “sheet” is space‑time, and the “rolling” is the motion of planets, stars, and light.

### Consequences

| Classic view | Relativistic view |
|--------------|-------------------|
| Gravity pulls objects with a force. | Objects follow the straightest possible paths (geodesics) in a curved space‑time. |
| Light is unaffected by gravity. | Light bends when it passes near massive objects (gravitational lensing). |
| Space and time are separate. | They’re intertwined into a single four‑dimensional “space‑time.” |

**Why it
```

---

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

**Response:**
```
<|channel|>analysis<|message|>Need explain democratic functions, institutions, processes.<|end|><|start|>assistant<|channel|>final<|message|>### 1.  What “democracy” means today  
Modern democracies are usually **constitutional republics** that combine:

| Core Feature | What it looks like in practice | Why it matters |
|--------------|--------------------------------|----------------|
| **Rule of law** | A written constitution, independent courts, and laws that apply to everyone, including leaders. | Prevents arbitrary power. |
| **Separation of powers** | Executive, legislative, and judicial branches that can check one another. | Avoids concentration of authority. |
| **Free and fair elections** | Regular, competitive elections with universal adult suffrage, secret ballots, and independent electoral commissions. | Gives citizens a direct voice. |
| **Political pluralism** | Multiple parties, interest groups, and civil society organizations. | Ensures diverse views are represented. |
| **Civil liberties** | Freedom of speech, press, assembly, and religion. | Enables public debate and accountability. |
| **Accountability mechanisms** | Transparency laws, audits, and media scrutiny. | Keeps officials answerable to the public. |

---

### 2.  How the system works day‑to‑day

| Stage | What happens | Key actors | Typical tools |
|-------|--------------|------------|---------------|
| **Policy agenda setting** | Citizens, NGOs, experts, and media highlight issues. | Civil society, think‑tanks, journalists | Public consultations, petitions, social media |
| **Legislative process** | Bills are drafted, debated, amended, and voted on. | Parliament/ congress, committees, lobbyists | Committee hearings, floor debates, voting records |
| **Executive action** | The head of state/government proposes policies, signs laws, and implements programs. | President/Prime Minister, ministries | Executive orders, budget proposals, regulatory agencies |
| **Judicial review** | Courts interpret laws and can strike down unconstitutional acts. | Supreme court, appellate courts | Judicial opinions, writs, injunctions |
| **Oversight & audit** | Independent bodies examine spending and performance. | Ombudsman, audit offices, watchdog NGOs | Audits, investigations, public reports |
| **Public participation** | Citizens vote, protest, or engage in deliberative forums. | Voters, activists, citizens | Elections, town halls, citizen assemblies |

---

### 3.  The “checks and balances” in action

| Branch | Typical checks on the
```

---

## 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)