Science GPT-OSS Model (24 Experts)

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

👥 Follow the Authors

Aman Priyanshu LinkedIn Twitter Website

Supriti Vijay LinkedIn Twitter Website

Introduction

This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 24 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 science 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.

Model Architecture & Statistics

Metric Value
Base Model openai/gpt-oss-20b
Architecture Mixture-of-Experts Transformer
Total Parameters ~16.1B (pruned from 21B)
Original Experts per Layer 32
Pruned Experts per Layer 24
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 24 experts per layer
  • No Retraining: Direct removal without additional training steps

Performance & Applications

Pruning Benefits

  • Smaller Memory Footprint: 75.0% 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 24 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 24 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

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the specialized model on CPU
model = AutoModelForCausalLM.from_pretrained(
    "AmanPriyanshu/gpt-oss-16.1b-specialized-science-pruned-moe-only-24-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-16.1b-specialized-science-pruned-moe-only-24-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

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-16.1b-specialized-science-pruned-moe-only-24-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-16.1b-specialized-science-pruned-moe-only-24-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:

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 concise explanation.<|end|><|start|>assistant<|channel|>final<|message|>**Photosynthesis in Plants – A Step‑by‑Step Overview**

| Stage | What Happens | Key Components | Result |
|-------|--------------|----------------|--------|
| **Light‑Dependent Reactions (Photophosphorylation)** | • Light energy is absorbed by chlorophyll and accessory pigments in the thylakoid membranes of chloroplasts. <br>• Excited electrons travel through the electron‑transport chain, pumping protons into the thylakoid lumen. <br>• The proton gradient powers ATP synthase to produce ATP. <br>• NADP⁺ is reduced to NADPH. | • Chlorophyll a & b, carotenoids, photosystems I & II, plastoquinone, cytochrome b₆f, ATP synthase, ferredoxin, NADP⁺ reductase. | • **ATP** (energy currency) and **NADPH** (reducing power) are generated. |
| **Calvin Cycle (Light‑Independent Reactions)** | • CO₂ is fixed into a 5‑carbon sugar (RuBP) by the enzyme **Rubisco**. <br>• The resulting 6‑carbon compound splits into two 3‑phosphoglycerate (3‑PGA) molecules. <br>• 3‑PGA is phosphorylated by ATP and reduced by NADPH to glyceraldehyde‑3‑phosphate (G3P). <br>• One G3P exits the cycle to contribute to glucose and other carbohydrates; the rest is recycled to regenerate RuBP. | • **Rubisco** (most abundant enzyme), **ATP**, **NADPH**, **RuBP**, **G3P**. | • **Glucose** (and other sugars) are synthesized; CO₂ is released back to the atmosphere. |

### Key Points

1. **Location** – All reactions occur in chloroplasts; light‑dependent reactions in thylakoid membranes, Calvin cycle in the stroma.
2. **Energy Flow** – Light energy → ATP + NADPH → CO₂ fixation → sugars.
3. **Water Splitting** – In photosystem II, water is oxidized, releasing O₂, electrons, and protons.
4. **Regulation** – Light intensity, CO₂ concentration, temperature, and stomatal opening control the overall rate.
5. **Outcome** – Plants convert atmospheric CO₂

Example 2

Question: What causes the greenhouse effect and how does it work?

Response:

<|channel|>analysis<|message|>Need explanation.<|end|><|start|>assistant<|channel|>final<|message|>### What Causes the Greenhouse Effect?

The greenhouse effect is a natural process that warms the Earth’s surface. It is caused by certain gases in the atmosphere—called **greenhouse gases (GHGs)**—that absorb and re‑emit infrared (IR) radiation. The main greenhouse gases are:

| Gas | Typical atmospheric concentration | Primary source |
|-----|-----------------------------------|----------------|
| **Water vapor (H₂O)** | ~0.3 % (variable) | Evaporation, transpiration |
| **Carbon dioxide (CO₂)** | ~0.04 % (≈ 410 ppm in 2024) | Fossil‑fuel combustion, deforestation, industrial processes |
| **Methane (CH₄)** | ~0.00018 % (≈ 1.9 ppm) | Agriculture (especially ruminants), wetlands, fossil‑fuel extraction |
| **Nitrous oxide (N₂O)** | ~0.00003 % (≈ 0.3 ppm) | Agriculture (fertilizers), combustion, industrial processes |
| **Ozone (O₃)** | ~0.00002 % (≈ 0.2 ppm) | Photochemical reactions in the troposphere |
| **CFCs and other halocarbons** | trace | Industrial refrigerants, aerosols (now largely phased out) |

These gases are **transparent to visible sunlight** (short‑wave radiation) but **opaque to the long‑wave infrared radiation** that the Earth emits after absorbing solar energy.

---

### How the Greenhouse Effect Works (Step‑by‑Step)

1. **Solar Radiation Reaches Earth**  
   - The Sun emits a spectrum of electromagnetic radiation, mostly in the visible and near‑infrared range (≈ 0.4–2.5 µm).  
   - The atmosphere is largely transparent to this short‑wave radiation, so most of it reaches the Earth’s surface.

2. **Earth Absorbs and Re‑Radiates Energy**  
   - The surface absorbs the incoming solar energy and warms.  
   - A warm surface emits thermal radiation (long‑wave IR) with a peak wavelength around 10 µm (according to Wien’s law).

3. **Greenhouse Gases Absorb IR Radiation**  
   - GHG molecules have vibrational modes that resonate with IR wavelengths.  
   - When a

Example 3

Question: Describe the structure and function of DNA.

Response:

<|channel|>analysis<|message|>Need concise description.<|end|><|start|>assistant<|channel|>final<|message|>**DNA (Deoxyribonucleic Acid)** is the hereditary material that stores and transmits genetic information in living organisms. Its structure and function can be understood in two complementary layers: the **molecular architecture** and the **biological roles** it plays.

---

## 1. Structural Features

| Feature | Description | Biological Significance |
|---------|-------------|------------------------|
| **Double‑Helix Backbone** | Two sugar‑phosphate chains wound around each other in a right‑handed helix. | Provides mechanical stability and protects the genetic code. |
| **Nucleotides** | Each unit consists of a deoxyribose sugar, a phosphate group, and a nitrogenous base. | The four bases (A, T, C, G) encode information. |
| **Base Pairing (Watson–Crick)** | Adenine (A) pairs with Thymine (T) via two hydrogen bonds; Cytosine (C) pairs with Guanine (G) via three hydrogen bonds. | Enables accurate replication and transcription. |
| **Antiparallel Orientation** | One strand runs 5′→3′, the other 3′→5′. | Essential for polymerase enzymes to read and synthesize DNA. |
| **Major & Minor Grooves** | Structural channels that allow protein binding. | Sites for transcription factors, polymerases, and repair enzymes. |
| **Chromatin Packaging** | DNA wraps around histone octamers to form nucleosomes → 30‑nm fiber → higher‑order structures. | Condensates DNA for mitosis, regulates gene accessibility. |
| **Topological Constraints** | Supercoiling, knotting, and catenation. | Managed by topoisomerases to facilitate replication and transcription. |

---

## 2. Functional Roles

| Function | Mechanism | Key Players |
|----------|-----------|------------|
| **Genetic Information Storage** | Sequence of bases encodes proteins, RNAs, regulatory elements. | Genes, introns, exons, regulatory sequences. |
| **Replication** | Semi‑conservative duplication; each new duplex contains one parental strand. | DNA polymerases (α, δ, ε), helicases, primase, ligase. |
| **Transcription** | RNA polymerase reads DNA template to synthesize mRNA. | RNA polymerase II (eukaryotes),

Citation

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}
}

References & Resources

Downloads last month
5
Safetensors
Model size
16.1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train AmanPriyanshu/gpt-oss-16.1b-specialized-science-pruned-moe-only-24-experts

Collection including AmanPriyanshu/gpt-oss-16.1b-specialized-science-pruned-moe-only-24-experts