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

Law 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 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	~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	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 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 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 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:

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 32 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-20.9b-specialized-law-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-law-pruned-moe-only-32-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-20.9b-specialized-law-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-law-pruned-moe-only-32-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 explain difference.<|end|><|start|>assistant<|channel|>final<|message|>**Civil law** and **criminal law** are two distinct branches of the legal system that serve different purposes, involve different parties, and have different procedures and consequences. Below is a concise comparison of the key differences:

| Feature | Civil Law | Criminal Law |
|---------|-----------|--------------|
| **Purpose** | To resolve disputes between private parties (individuals, businesses, organizations) and to compensate the injured party. | To protect society, maintain public order, and punish wrongdoing that violates criminal statutes. |
| **Parties Involved** | Plaintiff (the party bringing the claim) vs. Defendant (the party being sued). | State or government (prosecution) vs. Defendant (the accused). |
| **Burden of Proof** | “Preponderance of the evidence” – more likely than not (≈51%+). | “Beyond a reasonable doubt” – a very high standard of certainty. |
| **Typical Cases** | Contract disputes, property disputes, torts (e.g., negligence, defamation), family law (divorce, child custody), landlord‑tenant issues. | Theft, assault, murder, fraud, drug offenses, traffic violations that are criminal in nature. |
| **Outcome / Remedy** | Usually monetary damages, injunctions, specific performance, or declaratory judgments. | Penalties such as fines, probation, community service, imprisonment, or death (in jurisdictions that allow it). |
| **Procedure** | Civil litigation (court trial, arbitration, mediation). Parties often have more control over the process (e.g., can choose a mediator). | Criminal prosecution (court trial, plea bargaining). The state has the right to compel testimony and evidence; the defendant has constitutional protections (e.g., Miranda rights). |
| **Right to a Jury** | Jury trials are available but not required; many civil cases are decided by a judge. | Jury trials are a constitutional right in many jurisdictions for serious offenses (e.g., felony trials). |
| **Appeals** | Appeals are typically limited to legal errors; the standard is “de novo” review of the facts. | Appeals can be based on both legal errors and factual findings; higher courts often review the entire case. |
| **Time Limits** | Statutes of limitations restrict how long after an event a claim can be filed (e.g., 2–6 years for personal injury). | Generally no statute of limitations for serious crimes (e.g., murder

---

Example 2

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

Response:

<|channel|>analysis<|message|>Need explain due process.<|end|><|start|>assistant<|channel|>final<|message|>**Due process** is a foundational principle that guarantees fair treatment through the normal judicial system, especially as a citizen’s entitlement. It is a safeguard against arbitrary or unjust actions by the state, ensuring that legal proceedings are conducted with fairness, transparency, and respect for individual rights. The concept is embedded in many democratic legal systems, most notably in the U.S. Constitution (Fifth and Fourteenth Amendments) and in international human‑rights instruments such as the Universal Declaration of Human Rights and the International Covenant on Civil and Political Rights.

---

## 1. Historical Roots

| Era | Key Development | Significance |
|-----|-----------------|--------------|
| **Ancient Rome** | *Lex Aquilia* (c. 449 BC) – early codification of property rights and compensation. | Early recognition that the state must protect private property. |
| **English Common Law** | *Magna Carta* (1215) – “no man shall be imprisoned…without lawful judgment.” | Established the idea that the king must respect legal procedures. |
| **American Revolution** | *Bill of Rights* (1791) – Fifth and Fourteenth Amendments. | Codified due process as a constitutional guarantee. |
| **Modern Era** | *International Human Rights Law* (1948–present). | Universalized due process as a human right. |

---

## 2. Core Elements of Due Process

1. **Notice**  
   - The individual must be informed of the charges, claims, or actions against them.  
   - Includes the nature of the alleged wrongdoing, the evidence, and the potential consequences.

2. **Opportunity to Respond**  
   - The person must have a chance to present evidence, cross‑examine witnesses, and argue their case.  
   - This is often called the “right to a fair hearing” or “right to be heard.”

3. **Neutral Decision‑Maker**  
   - The adjudicator (judge, jury, or administrative body) must be impartial and free from bias.  
   - The decision must be based on the evidence and the law, not on arbitrary factors.

4. **Procedural Safeguards**  
   - Rules of evidence, procedural rules, and appellate review.  
   - These safeguards ensure consistency and prevent procedural abuses.

5. **Reasoned Decision**  
   - The outcome must be based on a clear, logical, and transparent rationale.  
   -

---

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.<|end|><|start|>assistant<|channel|>final<|message|>The First Amendment to the United States Constitution protects five core freedoms that are considered fundamental to a democratic society:

| # | Right | Key Elements |
|---|-------|--------------|
| 1 | **Freedom of Religion** | • **Establishment Clause** – the government may not establish an official religion or favor one religion over another.<br>• **Free‑Exercise Clause** – individuals may practice their religion without government interference. |
| 2 | **Freedom of Speech** | • Protects the right to express ideas, opinions, and information without censorship or punishment.<br>• Includes symbolic speech, political speech, and, in many cases, commercial speech (subject to certain restrictions). |
| 3 | **Freedom of the Press** | • Allows newspapers, broadcasters, and other media to report news and opinions without prior restraint by the government.<br>• Supports a free flow of information essential for an informed citizenry. |
| 4 | **Freedom of Assembly** | • Individuals may gather peacefully for protests, demonstrations, or other collective activities.<br>• The government may impose reasonable time, place, and manner restrictions but cannot prohibit assembly outright. |
| 5 | **Freedom to Petition the Government** | • Citizens can request changes to laws or policies, file complaints, or seek redress from the government without fear of retaliation. |

These five freedoms together form the backbone of civil liberties in the United States, ensuring that individuals can practice religion, express ideas, access information, gather for causes, and seek governmental accountability.<|return|>

---

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