AmanPriyanshu/gpt-oss-12.0b-specialized-law-pruned-moe-only-17-experts

Law GPT-OSS Model (17 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 17 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 law 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	~12.0B (pruned from 21B)
Original Experts per Layer	32
Pruned Experts per Layer	17
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	Law

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 law evaluation tasks
• Systematic Reduction: Reduced from 32 to 17 experts per layer
• No Retraining: Direct removal without additional training steps

Performance & Applications

Pruning Benefits

• Smaller Memory Footprint: 53.1% of original expert parameters
• Reduced Computational Load: Fewer routing decisions during inference
• Focused Capabilities: Retains experts relevant to law 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 legal domain model employs experts that demonstrated expertise during law-related tasks from MMLU legal subjects. These experts excel at legal reasoning, jurisprudence, and understanding of legal frameworks and procedures.

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 law tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 17 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 17 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-12.0b-specialized-law-pruned-moe-only-17-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-12.0b-specialized-law-pruned-moe-only-17-experts")

# Generate with the model
messages = [
    {"role": "user", "content": "What is the difference between civil and criminal law?"}
]

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-12.0b-specialized-law-pruned-moe-only-17-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-12.0b-specialized-law-pruned-moe-only-17-experts")

# Generate with the model
messages = [
    {"role": "user", "content": "What is the difference between civil and criminal law?"}
]

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 law tasks:

Example 1

Question: What is the difference between civil and criminal law?

Response:

<|channel|>analysis<|message|>Need concise explanation.<|end|><|start|>assistant<|channel|>final<|message|>**Civil law** and **criminal law** are the two main branches of the legal system, and they serve different purposes, involve different parties, and have different procedures and consequences.

| Feature | Civil Law | Criminal Law |
|---|---|---|
| **Purpose** | To resolve disputes between private parties (individuals, businesses, organizations). | To protect society and punish wrongdoing that harms the public or violates state authority. |
| **Parties** | Plaintiff (the person or entity that claims harm) vs. Defendant (the person or entity accused of wrongdoing). | State (prosecution) vs. Defendant (the accused). |
| **Burur of Proof** | “Pre‑pre‑balance” or “balance of probabilities” – the plaintiff must show the defendant is more likely than not to be liable. | “Beyond reasonable doubt” – the prosecution must prove the defendant’s guilt to a very high standard. |
| **Typical Cases** | Contract disputes, property claims, torts (negligence, defamation), family matters (divorce, child custody), estate matters. | Theft, assault, fraud, homicide, drug offenses, traffic violations, other offenses defined by statute. |
| **Remedies / Outcomes** | Usually monetary damages, injunctions, specific performance, or other equitable relief. | Penalties such as fines, probation, community service, imprisonment, or, in some jurisdictions, death penalty. |
| **Procedure** | Often starts with a civil lawsuit filed by the plaintiff; the defendant may file a counter‑claim. | Begins with a criminal complaint or indictment; the defendant is arrested or charged and then tried. |
| **Rights of the Defendant** | Right to a civil trial, to present evidence, to be represented by counsel (though not always required). | Constitutional rights: right to counsel, right to a jury (in many cases), right to remain silent, protection against self‑incrimination. |
| **Appeals** | Both parties can appeal, but the standard of review is usually “deference to the trial court’s findings.” | Appeals are typically limited to errors of law or procedure; the standard is “deference to the trial court’s findings.” |
| **Statistical Impact** | Civil cases are far more common and involve a wide range of everyday disputes. | Criminal cases are less frequent but carry higher social and legal stakes. |

### Key Take‑out

- **Civil law** is

---

Example 2

Question: Explain the concept of due process in legal systems.

Response:

<|channel|>analysis<|message|>Need explanation.<|end|><|start|>assistant<|channel|>final<|message|>### The Concept of Due Process (Due Process)

**Due process**—often called *due process* or *procedural due process*—is a fundamental principle in many legal systems that guarantees that the state (or any other party) must follow fair, lawful, and transparent procedures before it can deprate an individual of a protected interest (such as life, liberty, property, or a legal right). It is a safeguard against arbitrary, cap‑free, or abusive state action.

---

## 1.  What “Due Process” Means

| Term | Meaning |
|------|---------|
| **Due Process** | The requirement that the state must act in a fair, lawful, and transparent manner before deprating a person of a protected interest. |
| **Procedural Due Process** | The specific procedural steps (notice, hearing, decision, etc.) that must be followed. |
| **Substantive Due Process** | The requirement that the *content* of the decision (the law, policy, or action itself) must be lawful and not violate constitutional or statutory limits. |

In practice, *due process* is a two‑layer concept:

1. **Procedural Layer** – “How” the state acts (notice, hearing, right to counsel, etc.).
2. **Substantive Layer** – “What” the state does (the law or policy itself must be valid).

---

## 2.  Historical Roots

| Jurisdiction | Key Development |
|--------------|----------------|
| **United States** | *Due Process* is enunciated in the 4th, 5th, 6th, and 7th Amendments (e.g., “no person shall be deprived of life, liberty, or property … without due process of law”). |
| **United Kingdom** | The *Rule of Law* (e.g., *Rule 1* of the *Rule of Law* in *The Law of the Land*). |
| **Civil Law Countries** | *Due Process* is embedded in civil codes (e.g., Article 15 of the French Civil Code). |
| **International Law** | *Universal Declaration of Human Rights* (Article 12) and *International Covenant on Civil and Political Rights* (Article 12). |

---

## 3.  Core Elements of Procedural Due Process

| Element | Typical Requirements |
|----------|---------------------|
| **Notice** | The

---

Example 3

Question: What are the basic rights protected by the First Amendment?

Response:

<|channel|>analysis<|message|>Need to answer: freedom of religion, speech, press, assembly, petition, right to bear arms.<|end|><|start|>assistant<|channel|>final<|message|>The First Amendment to the United States Constitution protects a set of fundamental rights that are central to American civil liberties. These rights are often grouped into five main categories:

| **Right** | **Key Provisions** | **Typical Examples** |
|---|---|---|
| **Freedom of Religion** | • **Establishment Clause** – Congress may not establish a religion or favor one over another. <br>• **Free‑Relapse Clause** – Individuals may practice any religion or none. | • Religious worship in public and private spaces. <br>• Religious symbols on public property. |
| **Freedom of Speech** | • **Speech Clause** – Congress may not restrict the expression of ideas, opinions, or information. | • Political criticism, protest, satire, and commercial advertising. |
| **Freedom of the Press** | • **Press Clause** – Congress may not censor or control the press. | • Newspapers, radio, television, online media, and social‑media posts. |
| **Freedom of Assembly** | • **Assembly Clause** – Congress may not prohibit peaceful gathering of people. | • Peaceful marches, rallies, demonstrations, and public meetings. |
| **Right to Petition** | • **Petition Clause** – Congress may not deny the right of the people to petition the government. | • Writing to elected officials, filing petitions, and lobbying. |
| **Right to Bear Arms** | • **Second Amendment** (often grouped with the First Amendment in discussions of individual rights). | • Ownership and possession of firearms for lawful purposes. |

### How These Rights Are Interbalanced

- **Free‑Relapse vs. Establishment**: The First Amendment protects both the right to practice a religion and the right not to practice one, while preventing the state from endorsing or endorsing a particular religion.
- **Speech vs. Public Order**: The speech clause is broad, but courts recognize limits such as incun­tion, defamation, obsc‑ob, and “clear and present danger” tests.
- **Press vs. Privacy**: The press clause protects the press from government censorship, but it does not shield the press from defamation or privacy laws.
- **Assembly vs. Public Safety**: Peaceful assembly is protected, but the government can regulate time, place, and manner to maintain public safety.
- **Pet

---

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

• 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