README.md · AmanPriyanshu/gpt-oss-13.1b-specialized-all-pruned-moe-only-19-experts at main

gpt-oss-13.1b-specialized-all-pruned-moe-only-19-experts

File size: 18,193 Bytes

b1929e0

---
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 (19 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 19 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** | ~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** | 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 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 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 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:
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 19 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-13.1b-specialized-all-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-all-pruned-moe-only-19-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-13.1b-specialized-all-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-all-pruned-moe-only-19-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|>## 1.  What is Artificial Intelligence?

Artificial Intelligence (AI) is a field of computer science that seeks to build systems—usually software running on computers—that can perform tasks that, in humans, would normally require some form of *intelligence*.  These tasks include:

| Task | Human‑like behavior | Typical AI approach |
|------|---------------------|------------------------|
| Recognizing objects in pictures | “I see a cat” | Computer vision + deep learning |
| Understanding language | “What’s the weather?” | Natural‑language processing |
| Planning a route | “Take the shortest path” | Search algorithms, reinforcement learning |
| Playing chess | “I’ll move the queen to …” | Game‑theory + search |
| Diagnosing disease | “This looks like pneumonia” | Probabilistic inference, pattern matching |

In short, **AI is a set of algorithms and models that let machines learn from data, reason about it, and act in ways that mimic human cognition**.

---

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

AI systems are built from three core ingredients:

1. **Data** – The raw facts the system learns from (images, text, sensor readings, etc.).  
2. **Models / Algorithms** – The mathematical machinery that processes the data.  
3. **Training / Optimization** – The process of adjusting the model so it performs well on the task.

Below is a step‑by‑step outline of a typical AI pipeline:

| Step | What Happens | Key Techniques |
|------|--------------|----------------|
| **1. Problem Definition** | Decide what the AI should do (classification, generation, control, etc.). | Task specification, evaluation metrics |
| **2. Data Collection** | Gather labeled or unlabeled data that represents the problem. | Web scraping, sensor logs, crowdsourcing |
| **3. Pre‑processing** | Clean, normalize, and transform raw data into a usable format. | Feature extraction, tokenization, image augmentation |
| **4. Model Selection** | Pick a computational model that can learn from the data. | Decision trees, neural networks, probabilistic models |
| **5. Training** | Adjust the model’s internal parameters so it fits the data. | Gradient descent, back‑propagation, Bayesian inference |
| **6. Validation / Testing** | Measure how well the model general
```

---

### 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 – in plain words**

---

### 1. Two main ideas

| **What** | **Why it matters** |
|----------|---------------------|
| **The *speed of light is the same everywhere*** | Light (and any other thing that travels at that speed) can’t be used to tell you how fast you’re moving. |
| **“Time” and “space” are one thing, not separate** | How far apart things are and how long they take to get there are linked. If you’re moving fast, the way you measure distance and time changes. |

---

### 2. The “speed of light is the same everywhere”

- Imagine you’re on a perfectly straight, super‑fast train. Inside the train, you can’t feel the motion; everything looks normal.
- If you shine a flash of light from the front of the train, the light travels at the same speed whether it’s inside the train or outside in the world.
- Because the light’s speed never changes, it becomes a universal “clock” and “yardstick.”  
  *If you measure how long a light takes to travel a known distance, you can deduce how fast you’re moving.*

---

### 3. “Space + Time = One Thing”

- Think of a movie frame: the picture (space) and the motion of the picture (time) are together.  
- In relativity, if you’re moving very fast, the *order* of events can change.  
  *Example:* Two clocks, one on Earth, one on a spaceship zooming at 90% the speed of light.  
  - To the Earth clock, the spaceship clock ticks slower.  
  - To the spaceship clock, the Earth clock ticks slower.  
  Both are right; the difference comes from the fact that “time” depends on how you’re moving through space.

---

### 4. Everyday consequences (tiny, but real)

| **Effect** | **What you notice** |
|------------|---------------------|
| **Gravitational pull** | The heavier a planet, the stronger the pull. |
| **Light bending** | Light from a star can be bent by a massive planet or black hole. |
| **Time dilation** | Astronauts on a very fast spacecraft age a fraction slower than people on Earth. |

---

### 5. Bottom‑line

- **Speed
```

---

### 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|>**Democracy in the modern world is not a single, one‑style system but a family of institutions, practices, and norms that together create a “government by the people, for the people.”**  
Below is a practical map of how those pieces fit together in most contemporary democratic states.

| Key Element | What It Looks Like Today | Why It Matters | Typical Variations |
|-------------|---------------------------|----------------|---------------------|
| **Constitution / Charter** | Written or unwritten set of rules that defines the state’s structure, powers, and limits. | Provides a legal framework that protects rights and limits government power. | Some countries have a single codified constitution (e.g., the U.S.); others rely on a mix of statutes, case law, and conventions (e.g., the U.K.). |
| **Separation of Powers** | Executive (government), Legislature (law‑making), Judiciary (law‑interpreting). | Prevents concentration of power; each branch can check the others. | In some systems the legislature is bicameral (two houses) or unicameral; the judiciary may be independent or subject to political appointments. |
| **Rule of Law** | All citizens, including leaders, are subject to the same laws. | Ensures fairness, accountability, and predictability. | Some democrates have “rule of law” but still allow for executive decrements in emergencies (e.g., emergency powers). |
| **Free, Fair, and Regular Elections** | Universal adult suff voting, secret ballots, independent electoral commissions. | The core mechanism for choosing representatives and, in some cases, the executive. | Variations: direct presidential elections, parliamentary votes, referendums, proportional vs. majoritarian systems. |
| **Political Pluralism & Competition** | Multiple parties, independent media, civil‑society groups. | Allows diverse views to compete and be represented. | Some democrates have a two‑party system; others have many parties with coalition governments. |
| **Civil‑Society Rights** | Freedom of expression, assembly, association, religion, and privacy. | Enables citizens to hold power accountable and to advocate for change. | The scope and enforcement of these rights vary; some democrates have strong protections, others have restrictions (e.g., on hate speech, national security). |
| **Independent Judiciary** | Courts that can review laws, executive actions, and protect
```

---

## 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)