AmanPriyanshu/gpt-oss-17.3b-specialized-law-pruned-moe-only-26-experts

Law GPT-OSS Model (26 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 26 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	~17.3B (pruned from 21B)
Original Experts per Layer	32
Pruned Experts per Layer	26
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 26 experts per layer
• No Retraining: Direct removal without additional training steps

Performance & Applications

Pruning Benefits

• Smaller Memory Footprint: 81.2% 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 26 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 26 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-17.3b-specialized-law-pruned-moe-only-26-experts", 
    torch_dtype=torch.bfloat16, 
    device_map="cpu", 
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-17.3b-specialized-law-pruned-moe-only-26-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-17.3b-specialized-law-pruned-moe-only-26-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-17.3b-specialized-law-pruned-moe-only-26-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 outcomes. Below is a concise comparison that highlights the key differences:

| Feature | Civil Law | Criminal Law |
|---------|-----------|--------------|
| **Purpose** | To resolve disputes between private parties (individuals, businesses, organizations). | To protect society by punishing conduct that is deemed harmful or dangerous to the public. |
| **Parties** | 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 (≈99% certainty). |
| **Outcome / Remedy** | Usually monetary damages, injunctions, specific performance, or declaratory judgments. | Penalties such as fines, imprisonment, probation, community service, or other sanctions. |
| **Nature of Wrong** | Wrongful acts that cause loss or injury (e.g., breach of contract, negligence, defamation). | Acts that violate criminal statutes (e.g., theft, assault, fraud, homicide). |
| **Procedure** | Civil courts (or arbitration/mediation) handle the case; proceedings are often less formal and can be initiated by the plaintiff. | Criminal courts (or special criminal tribunals) handle the case; the prosecution initiates the action, and the process is more formal with strict procedural safeguards. |
| **Rights of the Defendant** | Right to a fair trial, but the defendant is not presumed guilty. | Presumption of innocence; the defendant must be proven guilty. |
| **Appeal** | Both parties can appeal, but the standard of review may differ. | Appeals are typically limited to procedural or substantive errors; the appellate court reviews the trial court’s application of law. |
| **Timing** | Statutes of limitations often apply (e.g., 3–10 years depending on the claim). | Statutes of limitations are usually longer or may not apply to serious crimes (e.g., murder). |
| **Examples** | Contract disputes, property disputes, tort claims, family law matters (divorce, child custody). | Theft, assault, drug offenses, fraud, homicide. |

###

---

Example 2

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

Response:

<|channel|>analysis<|message|>Need explain concept, origins, procedural and substantive, examples.<|end|><|start|>assistant<|channel|>final<|message|>**Due Process – A Cornerstone of Fair Legal Systems**

---

### 1. What is Due Process?

**Due process** is a constitutional or statutory guarantee that the state will not deprive an individual of life, liberty, or property without a fair and orderly legal procedure. It is a two‑fold doctrine:

| **Component** | **Meaning** | **Typical Manifestations** |
|---------------|-------------|----------------------------|
| **Procedural Due Process** | The *process* that must be followed before the government can act. | Notice of charges, right to be heard, impartial judge, opportunity to present evidence, right to counsel, record‑keeping, appeal rights. |
| **Substantive Due Process** | The *content* of the law itself must be fair, not arbitrary, and not violate fundamental rights. | Laws that are overly vague, discriminatory, or that infringe on fundamental liberties (e.g., freedom of speech, privacy). |

---

### 2. Historical Roots

| **Era** | **Key Milestones** | **Impact on Due Process** |
|---------|--------------------|---------------------------|
| **Ancient** | Roman *Lex Aquila* (c. 200 BCE) – “no one may be punished without a trial.” | Early recognition that punishment requires a trial. |
| **Medieval** | Magna Carta (1215) – “no free man shall be imprisoned…without lawful judgment.” | Established the principle that the king must respect legal procedures. |
| **American** | U.S. Constitution, 5th & 14th Amendments (1791, 1868) – “No person shall...be deprived of life, liberty, or property...without due process of law.” | Codified procedural and substantive due process in the U.S. legal system. |
| **Modern** | European Convention on Human Rights (1950), International Covenant on Civil and Political Rights (1960) | Internationally recognized due process as a human right. |

---

### 3. Procedural Due Process – “How” the Law is Applied

| **Step** | **Explanation** | **Typical Legal Requirement** |
|----------|-----------------|------------------------------|
| **Notice** | The person must be informed of the action or charge. | Written notice of arrest, indictment, or administrative decision. |
| **Opportunity to Respond** |

---

Example 3

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

Response:

<|channel|>analysis<|message|>Need to answer: basic rights: freedom of religion, speech, press, assembly, petition. Provide explanation.<|end|><|start|>assistant<|channel|>final<|message|>**The First Amendment protects five core freedoms** (often called the “Five Freedoms”):

| Freedom | What it protects | Typical examples |
|---------|------------------|------------------|
| **Religion** | 1. **Free exercise** – the right to practice any religion or none at all. 2. **Establishment** – the government may not create or favor a religion. | Churches, mosques, temples, or a person’s choice to be non‑religious. |
| **Speech** | The right to express ideas, opinions, and information without government censorship. | Political speeches, protests, social‑media posts, artistic performances. |
| **Press** | Freedom of the press to publish news, opinions, and investigative reporting. | Newspapers, magazines, online blogs, television news. |
| **Assembly** | The right to gather peacefully for protest, demonstration, or other collective action. | Marches, rallies, sit‑ins, town‑hall meetings. |
| **Petition** | The right to petition the government for redress of grievances. | Letter‑writing campaigns, online petitions, lobbying. |

### How they work together

- **Freedom of religion** ensures that the state cannot impose a religion or interfere with religious practice.
- **Freedom of speech** and **press** give individuals the ability to share ideas and hold the government accountable.
- **Freedom of assembly** lets people come together to express those ideas collectively.
- **Freedom of petition** provides a formal channel to ask the government to change policies or address wrongs.

These rights are not absolute; courts allow certain restrictions (e.g., for national security, public safety, or to prevent defamation). However, the First Amendment sets the baseline that the government may not unduly limit these freedoms.<|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