---
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
  - science
  - specialized
  - efficient
  - transformer
  - causal-lm
  - text-generation
  - pytorch
  - pruned-model
  - domain-specific
---

# Science GPT-OSS Model (12 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 12 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 science 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** | ~9.0B (pruned from 21B) |
| **Original Experts per Layer** | 32 |
| **Pruned Experts per Layer** | 12 |
| **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** | Science |

## 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 science evaluation tasks
- **Systematic Reduction**: Reduced from 32 to 12 experts per layer
- **No Retraining**: Direct removal without additional training steps

## Performance & Applications

### Pruning Benefits
- **Smaller Memory Footprint**: 37.5% of original expert parameters
- **Reduced Computational Load**: Fewer routing decisions during inference
- **Focused Capabilities**: Retains experts relevant to science 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 science-specialized model leverages experts that showed high activation patterns during scientific reasoning tasks from GPQA (physics, chemistry, biology) and MMLU science domains. These experts demonstrate superior performance on complex scientific problem-solving and technical knowledge recall.

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 science tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 12 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 12 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-9.0b-specialized-science-pruned-moe-only-12-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-9.0b-specialized-science-pruned-moe-only-12-experts")

# Generate with the model
messages = [
    {"role": "user", "content": "Explain the process of photosynthesis in plants."}
]

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-9.0b-specialized-science-pruned-moe-only-12-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-9.0b-specialized-science-pruned-moe-only-12-experts")

# Generate with the model
messages = [
    {"role": "user", "content": "Explain the process of photosynthesis in plants."}
]

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 science tasks:

### Example 1
**Question:** Explain the process of photosynthesis in plants.

**Response:**
```
<|channel|>analysis<|message|>Need to explain photosynthesis.<|end|><|start|>assistant<|channel|>final<|message|>**Photosynthesis – the life‑fueling process of plants**

---

### 1.  What is photosynthesis?

Photosynthesis is the conversion of light energy into chemical energy that plants use to grow, reproduce and maintain their life.  In short, plants “burn” sunlight to make a new molecule of glucose (the sweet sugar that fuels life).

---

### 2.  The key parts of the process

| **Subcellular organ** | **Molecule** | **Function** |
|---|---|---|
| **Chloroplast** | – | The site where the reaction happens |
| **Stromule** | – | The tiny “poore” that lets light reach the chlorophyll |
| **Chlorophyll** (main pigment) | – | Abs absorbs light, especially red‑green wavelengths |
| **Car‑t‑p** | – | The enzyme that turns the light‑energy into chemical energy |
| **Glucose** | – | The product that fuels the plant’s metabolism |

---

### 3.  Step‑by‑step

1. **Light enters the leaf**  
   Sunlight (or artificial light) penetrates the leaf surface.  The chlorophyll molecules inside the chloroplast absorb photons of the red‑green part of the spectrum (≈400–700 nm).

2. **Photon energy is captured**  
   Each photon carries a certain amount of energy.  When a chlorophyll molecule absorbs a photon, the energy is transferred to the excited state of the *excated* electron.

3. **Excited electron moves**  
   The excited electron is transferred to the *excited* state of the *car‑t‑p* enzyme (the “red‑green” reaction centre).  The enzyme has two sub‑units:  
   * **P** (phph‑ph‑ph) – the “ph” part that accepts the photon.  
   * **N** (nit‑nit‑nit) – the “nit” part that accepts the electron.  

   The electron moves from P to N, creating a “chemical” potential difference (a red‑green charge).

4. **Red‑green charge drives the reaction**  
   The red‑green charge is used to drive the reaction that splits two molecules of water (H₂O) into one molecule of oxygen (O₂) and one molecule of hydrogen (H₂).  The
```

---

### Example 2
**Question:** What causes the greenhouse effect and how does it work?

**Response:**
```
<|channel|>analysis<|message|>Need explain greenhouse effect causes and mechanism.<|end|><|start|>assistant<|channel|>final<|message|>**The “greenhouse effect” (or simply the greenhouse effect) is the phenomenon that the Earth’s atmosphere traps heat, keeping the surface warmer than it would be if the atmosphere were transparent to all radiation.**  

---

## 1.  What causes it?

| **Key factors** | **How they contribute** |
|---|---|
| **Atmospheric composition** | The atmosphere contains gases that are very good at absorbing and re‑emtrating infrared (IR) radiation. The main gases are:  
  * **Water vapor (H₂O)** – abundant in the lower tro‑m, highly absorbing in the IR.  
  * **Carbon dioxide (CO₂)** – a strong greenhouse gas, especially in the 10–15 µm window.  
  * **Methane (CH₄) and nit‑ox (N₂O)** – also absorb IR. | These gases absorb incoming IR from the Earth’s surface, heat them up, and then re‑emit the same IR back to space. |
| **O‑zone and aerosols** | O‑zone absorbs UV, but also scatters visible light. Aerosols (dust, clouds, soot) scatter and absorb visible/IR light, adding a “back‑reflection” that warms the surface. | |
| **Radiative imbalance** | The Earth receives more solar radiation than it emits. The greenhouse effect is the *feedback* that keeps the surface warmer, which in turn increases the amount of greenhouse gases (e.g., more water vapor, more CO₂). | |

The main driver is the **absorption–re‑emission cycle** of the greenhouse gases.  

---

## 2.  How it works – the radiative transfer picture

1. **Solar radiation enters the atmosphere**  
   * Short‑wave (UV, visible) is largely absorbed by the upper‑tro‑m and reflected by clouds.  
   * Long‑wave (IR) penetrates deeper, reaching the surface.

2. **Surface emits IR**  
   The Earth’s surface emits a black‑body spectrum (≈ 10 µm peak). This IR travels upward through the atmosphere.

3. **Absorption by greenhouse gases**  
   As the IR passes, gases like CO₂, H₂O, CH₄ absorb specific wavelengths. The energy is transferred to the gas molecules, raising their temperature.

4. **Re
```

---

### Example 3
**Question:** Describe the structure and function of DNA.

**Response:**
```
<|channel|>analysis<|message|>Need to explain DNA structure: double helix, 5 strands, 10 base pairs, etc. Function: genetic information, coding, replication, transcription, etc.<|end|><|start|>assistant<|channel|>final<|message|>**DNA: the structure and its function**

---

## 1.  The DNA molecule – a double‑helix of 5 strands

| **Key features** | **What does it do?** |
|---|---|
| **Double‑helix** – two strands wind around each other, giving a “hand‑crossed” shape. | Provides a compact, stable structure that protects the genetic information. |
| **5 strands** – 5 parallel strands of the sugar backbone (phosphodiol). | Gives a “hand‑crossed” arrangement that allows the base‑pairs to be read from one side. |
| **10 base‑pairs** – 5 complementary base‑pairs (A–T, G–C) on each strand. | Each base‑pair is a “code” that tells the protein what it should do. |
| **Sugar backbone** – de‑phosphoryated sugar (d‑de‑phosphoryated). | Provides a scaffold that keeps the DNA stable and ready for replication. |
| **Base‑pairs** – 5 complementary pairs (A–T, G–C). | The complementary pairs are the “code” that can be read by the enzymes that will read the DNA. |
| **Base‑pairs** – 5 complementary pairs (A–T, G–C). | The complementary pairs are the “code” that can be read by the enzymes that will read the DNA. |
| **Base‑pairs** – 5 complementary pairs (A–T, G–C). | The complementary pairs are the “code” that can be read by the enzymes that will read the DNA. |
| **Base‑pairs** – 5 complementary pairs (A–T, G–C). | The complementary pairs are the “code” that can be read by the enzymes that will read the DNA. |
| **Base‑pairs** – 5 complementary pairs (A–T, G–C). | The complementary pairs are the “code” that can be read by the enzymes that will read the DNA. |
| **Base‑pairs** – 5 complementary pairs (A–T, G–C). | The complementary pairs are the “code” that can be read by the enzymes that will
```

---

## 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)