You can deploy this model and start earning money today!

AmanPriyanshu/
gpt-oss-20.9b-specialized-science-pruned-moe-only-32-experts

Name: gpt-oss-20.9b-specialized-science-pruned-moe-only-32-experts
Availability: InStock

Views

No views yet

Science GPT-OSS Model (32 Experts)

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

👥 Follow the Authors

Aman Priyanshu

Supriti Vijay

Introduction

This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 32 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	~20.9B (pruned from 21B)
Original Experts per Layer	32
Pruned Experts per Layer	32
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:

Analyzing Usage Patterns: Tracking which experts activate most frequently for specific tasks
Removing Underutilized Experts: Discarding experts with low activation rates for the target domain
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 32 experts per layer
No Retraining: Direct removal without additional training steps

Performance & Applications

Pruning Benefits

Smaller Memory Footprint: 100.0% 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 32 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:

Activation Analysis: Comprehensive evaluation of expert usage patterns across domain-specific tasks
Expert Ranking: Identification of the most frequently activated experts for target domains
Systematic Pruning: Reduction from 32 to 32 experts while preserving router functionality
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-20.9b-specialized-science-pruned-moe-only-32-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-20.9b-specialized-science-pruned-moe-only-32-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-20.9b-specialized-science-pruned-moe-only-32-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-20.9b-specialized-science-pruned-moe-only-32-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 concise explanation.<|end|><|start|>assistant<|channel|>final<|message|>**Photosynthesis** is the process by which green plants, algae, and some bacteria convert light energy into chemical energy, producing sugars that fuel their growth and releasing oxygen as a by‑product. It takes place mainly in the chloroplasts of plant cells, specifically in the thylakoid membranes where chlorophyll pigments capture light.

---

### 1. Light‑Dependent Reactions (Thylakoid Membrane)

| Step | What Happens | Key Components |
|------|--------------|----------------|
| **Light Capture** | Chlorophyll a and accessory pigments absorb photons. | Chlorophyll a, b, carotenoids |
| **Water Splitting (Photolysis)** | Light energy splits H₂O into O₂, H⁺, and electrons. | Photosystem II (PSII) |
| **Electron Transport Chain (ETC)** | Electrons move through PSII → plastoquinone → cytochrome b₆f → plastocyanin → Photosystem I (PSI). | PSII, cytochrome b₆f, PSI |
| **ATP Formation** | Proton gradient drives ATP synthase. | ATP synthase |
| **NADPH Production** | Electrons reduce NADP⁺ to NADPH. | Ferredoxin, NADP⁺ reductase |
| **Oxygen Release** | O₂ is released into the atmosphere. | PSII |

**Result:** 2 ATP, 2 NADPH, and 1 O₂ per cycle of light absorption.

---

### 2. Light‑Independent Reactions (Calvin‑Benson Cycle, Stroma)

| Step | What Happens | Key Enzymes |
|------|--------------|-------------|
| **Carbon Fixation** | CO₂ is attached to ribulose‑1,5‑bisphosphate (RuBP) by Rubisco, forming 3‑phosphoglycerate (3‑PGA). | Rubisco |
| **Reduction Phase** | 3‑PGA is phosphorylated by ATP and reduced by NADPH to glyceraldehyde‑3‑phosphate (G3P). | ATP, NADPH |
| **Regeneration of RuBP** | Some G3P molecules are used to regenerate RuBP, allowing the cycle to continue. | Various enzymes (e.g., phosphoglycerate kinase, glyceraldehyde‑3‑phosphate dehydrogenase) |
|

Example 2

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

Response:

<|channel|>analysis<|message|>Need explain greenhouse effect causes, greenhouse gases, absorption, re-emission, energy balance.<|end|><|start|>assistant<|channel|>final<|message|>### The Greenhouse Effect in a Nutshell

The **greenhouse effect** is the natural process that keeps Earth’s surface warm enough to support life. It works because certain gases in the atmosphere (the *greenhouse gases*, or GHGs) trap heat that would otherwise escape into space. Think of it as a blanket that keeps the planet from cooling too quickly.

---

## 1. What Causes the Greenhouse Effect?

| Cause | How It Happens |
|-------|----------------|
| **Solar Radiation** | The Sun emits energy across a spectrum of wavelengths. Short‑wave (visible and UV) radiation passes through the atmosphere largely unimpeded and warms the Earth’s surface. |
| **Atmospheric Gases** | Gases such as **water vapor (H₂O)**, **carbon dioxide (CO₂)**, **methane (CH₄)**, **nitrous oxide (N₂O)**, and **ozone (O₃)** absorb and re‑emit infrared (IR) radiation. |
| **Surface Emission** | The warmed surface emits IR radiation back toward space. |
| **Re‑emission by GHGs** | Greenhouse gases absorb some of that IR and then re‑emit it in all directions, including back toward the surface, adding extra warmth. |

The net result is that the surface temperature is higher than it would be if the atmosphere were transparent to IR radiation.

---

## 2. How Does It Work? (Step‑by‑Step)

1. **Solar Energy Arrives**  
   - The Sun’s energy is mostly in the visible spectrum (≈ 50–70 % of total solar irradiance).  
   - About 30 % is reflected by clouds, ice, and the atmosphere (albedo).  
   - Roughly 70 % reaches Earth’s surface.

2. **Surface Absorbs and Re‑emits Energy**  
   - The surface absorbs the incoming short‑wave radiation and warms.  
   - A warmed surface emits long‑wave (infrared) radiation, with a peak wavelength around 10 µm (depending on temperature).

3. **Greenhouse Gases Absorb IR**  
   - GHG molecules have vibrational modes that resonate with IR wavelengths.  
   - When IR photons hit a GHG molecule, the molecule absorbs the energy, raising

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 carries the genetic blueprint for all living organisms. Its structure and function can be understood in two complementary parts:

---

## 1. Structure

| Feature | Description | Biological Significance |
|---------|-------------|------------------------|
| **Double‑helix** | Two antiparallel strands wound around each other like a twisted ladder. | Provides stability and a compact way to store long sequences of genetic information. |
| **Backbone** | Alternating deoxyribose sugars and phosphate groups. | Forms a sturdy, negatively charged scaffold that protects the bases and allows enzymes to read the sequence. |
| **Nucleobases** | Four nitrogenous bases: Adenine (A), Thymine (T), Cytosine (C), Guanine (G). | Encode genetic information through base‑pairing (A‑T, C‑G). |
| **Base‑pairing** | Hydrogen bonds link complementary bases across the two strands. | Ensures accurate replication and transcription. |
| **Major & Minor grooves** | Unequal spaces between the strands. | Provide binding sites for proteins (e.g., transcription factors, polymerases). |
| **Supercoiling & Chromatin** | In eukaryotes, DNA is wrapped around histone proteins forming nucleosomes, which further coil into higher‑order structures. | Enables compaction of millions of base pairs into the nucleus while regulating access to genes. |

---

## 2. Function

| Function | How It Works | Key Players |
|----------|--------------|-------------|
| **Genetic Information Storage** | The linear sequence of bases encodes instructions for building proteins and RNA molecules. | Genes, regulatory elements |
| **Replication** | DNA polymerases synthesize a new complementary strand, producing two identical DNA molecules. | DNA polymerase, helicase, ligase |
| **Transcription** | RNA polymerase reads a DNA template strand and synthesizes messenger RNA (mRNA). | RNA polymerase, transcription factors |
| **Translation (Protein Synthesis)** | Ribosomes read mRNA codons and assemble amino acids into polypeptide chains. | tRNA, ribosomes, aminoacyl‑tRNA synthetases |
| **Mutation & Evolution** | Errors in replication or damage can change base sequences, creating genetic diversity. | DNA repair enzymes, mutagens |
| **Regulation of

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
Model Hub: GPT-OSS-20B on Hugging Face
Expert Analysis Dataset: GPT-OSS-20B MoE Expert Activations
Project Page: GPT-OSS MoE Expert Fingerprinting
GitHub Repository: OpenAI GPT-OSS