You can deploy this model and start earning money today!

AmanPriyanshu/
gpt-oss-13.1b-specialized-science-pruned-moe-only-19-experts

Name: gpt-oss-13.1b-specialized-science-pruned-moe-only-19-experts
Availability: InStock

Views

No views yet

Science GPT-OSS Model (19 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 19 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	~13.1B (pruned from 21B)
Original Experts per Layer	32
Pruned Experts per Layer	19
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 19 experts per layer
No Retraining: Direct removal without additional training steps

Performance & Applications

Pruning Benefits

Smaller Memory Footprint: 59.4% 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 19 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 19 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-13.1b-specialized-science-pruned-moe-only-19-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-13.1b-specialized-science-pruned-moe-only-19-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-13.1b-specialized-science-pruned-moe-only-19-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-13.1b-specialized-science-pruned-moe-only-19-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 in Plants – A Step‑by‑Step Overview**

| **Stage** | **What Happens** | **Key Players** | **Result** |
|---|---|---|---|
| 1. **Light Capture** | Sunlight hits chlorophyll‑containing pigments in the thylakoid membranes of chloroplasts. | • Chlorophyll a & b<br>• Carotenoids (protective pigments) | • Excited electrons are released from chlorophyll |
| 2. **Electron Transport Chain (ETC)** | Excited electrons travel through a series of proteins (photosystem II → cytochrome b₆₀ → photosystem I → cytochrome a₁₀). | • Photosystem II (PSII)<br>• Photosystem I (PSI)<br>• Cytochrome complexes | • Energy is stored in ATP (via proton gradient) and NADPH (via electron reduction) |
| 3. **Carbon Fixation (Calvin Cycle)** | CO₂ from the air is incorporated into sugars using ATP and NADPH. | • Ribulose‑1‑phosphate carboxylase‑oxygenase (RuBisCO)<br>• Other enzymes (e.g., GAPDH, PGK) | • Glyceraldehyde‑3‑phosphate → 3‑phosphoglycerate → 1‑bromopentose → glucose & other carbohydrates |
| 4. **Oxygen Release** | Water molecules split (photolysis) in PSII, releasing O₂ as a by‑product. | • Water‑splitting enzyme (H₂O → 2H⁺ + 4e⁻ + O₂) | • O₂ diffuses out through stomata |

### Quick‑Line Summary

1. **Light energy** excites chlorophyll → electrons are freed.  
2. Electrons travel through the ETC, generating **ATP** and **NADPH**.  
3. **ATP + NADPH** power the Calvin cycle, fixing atmospheric CO₂ into glucose.  
4. **Water** is split to supply electrons, releasing **oxygen**.

**Result:** Plants convert light, CO₂, and H₂O into glucose (energy source) and O₂ (life‑supporting gas).<|return|>

Example 2

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

Response:

<|channel|>analysis<|message|>Need explain greenhouse effect, causes, mechanism.<|end|><|start|>assistant<|channel|>final<|message|>### 1. What is the Greenhouse Effect?

The **greenhouse effect** is a natural process that keeps Earth’s surface warm enough to support life. It works by trapping a portion of the Sun’s energy that would otherwise escape into space, much like a greenhouse traps heat.

---

## 2. What Causes the Greenhouse Effect?

| **Primary Cause** | **How it Works** |
|-------------------|-------------------|
| **Atmospheric Greenhouse Gases** | Gases that absorb and re‑radiate infrared (IR) radiation emitted by the Earth’s surface. |
| **Key Gases** | • **Water vapor (H₂O)** – ~70 % of the natural greenhouse effect. <br>• **Carbon dioxide (CO₂)** – ~20 % (human‑produced ~0.04 % of the atmosphere). <br>• **Methane (CH₄)** – ~1 % (human‑produced ~0.0002 %). <br>• **Nitrous oxide (N₂O)**, **ozone (O₃)**, and **CFCs** – smaller but potent contributors. |
| **Human Activities** | • Burning fossil fuels (coal, oil, gas). <br>• Deforection and land‑use changes. <br>• Agriculture (especially livestock, rice paddies). <br>• Industrial processes (cement, ammonia). |
| **Resulting Increase** | Since the pre‑industrial era (~1850), atmospheric CO₂ has risen from ~280 ppm to ~420 ppm, and the total greenhouse‑gas forcing has increased by ~1.8 W m⁻². |

---

## 3. How Does the Greenhouse Effect Work? (Step‑by‑Step)

1. **Solar Radiation Arranges**  
   * The Sun emits mainly visible and near‑infrared (NIR) light.  
   * About 30 % of this is reflected by clouds, ice, and the Earth’s surface.  
   * Roughly 70 % reaches the surface and is absorbed, warming the land, oceans, and atmosphere.

2. **Earth’s Surface Re‑Radiates**  
   * The warmed surface emits thermal radiation, mostly in the **long‑wave (infra‑red)** part of the spectrum (≈ 5–50

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 genetic material that carries the instructions for building and regulating all living organisms. Its structure and function can be broken down into several key components:

| **Structure** | **Function** |
|---|---|
| **Double‑Helix Backbone** – Two complementary strands of nucleotides wound around each other in a right‑handed helix. The sugar‑phosphate backbone provides structural stability and protects the genetic code. | **Genetic Information Storage** – The sequence of nucleotides encodes the instructions for proteins and RNA molecules. |
| **Nucleotides** – Four bases (Adenine, Thymine, Cytosine, Guanine) pair via hydrogen bonds (A‑T, C‑G). | **Protein Synthesis** – During transcription, RNA polymerase reads the DNA sequence and produces messenger RNA (mRNA), which is translated into proteins. |
| **Major & Minor Groes** – The helical twist creates grooves that allow proteins (e.g., transcription factors, nucleases) to access the DNA. | **Regulation of Gene Expression** – DNA‑binding proteins recognize specific sequences in the grooves to activate or repress genes. |
| **Chromatin Packaging** – In eukaryotes, DNA wraps around histone proteins to form nucleosomes, which further coil into higher‑order structures (chromosomes). | **Cell Division & Chromosome Segregation** – Proper packaging ensures accurate duplication and segregation of genetic material during mitosis and meiosis. |
| **Replication Origin Sites** – Specific sequences (e.g., AT‑rich “origin of replication”) where DNA helicase unwinds the strands. | **DNA Replication** – The double helix is copied during cell division, ensuring each daughter cell receives a complete genome. |
| **Repair Mechanisms** – Enzymes (e.g., DNA polymerase, ligase, helicase) correct mismatches, breaks, and mutations. | **Genome Integrity** – Maintains fidelity of genetic information across generations. |

**Key Functional Roles**

1. **Encoding Genetic Code** – Determines amino acid sequences of proteins, regulatory RNAs, and other biomolecules.  
2. **Regulation of Gene Activity** – DNA sequences (promoters, enhancers, silencers) control when and where genes are expressed.  
3. **Inheritance** – Passed from parent to offspring,

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