Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 12 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 safety 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.
| Metric | Value |
|---|---|
| Base Model | openai/gpt-oss-20b |
| Architecture | Mixture-of-Experts Transformer |
| Total Parameters | ~9.0B (pruned from 21B) |
| Original Experts per Layer | 32 |
| Pruned Experts per Layer | 12 |
| 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 | Safety |
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:
Note: Performance may vary depending on how well the pruned experts match your specific use case.
This safety-focused model uses experts that performed well on safety evaluation tasks from SORRY-Bench. These experts are specialized in identifying and appropriately responding to potentially harmful content while maintaining helpful capabilities.
The expert selection process utilized our comprehensive analysis of router activation patterns across multiple evaluation benchmarks:
By identifying experts that consistently activated for safety tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 12 experts per layer.
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.
Our approach involves:
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.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# Load the specialized model on CPU
model = AutoModelForCausalLM.from_pretrained(
"AmanPriyanshu/gpt-oss-9.0b-specialized-safety-pruned-moe-only-12-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-9.0b-specialized-safety-pruned-moe-only-12-experts")
# Generate with the model
messages = [
{"role": "user", "content": "What should someone do if they encounter cyberbullying online?"}
]
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)
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-9.0b-specialized-safety-pruned-moe-only-12-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-9.0b-specialized-safety-pruned-moe-only-12-experts")
# Generate with the model
messages = [
{"role": "user", "content": "What should someone do if they encounter cyberbullying online?"}
]
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)
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
The following examples demonstrate the model's performance on safety tasks:
Question: What should someone do if they encounter cyberbullying online?
Response:
<|channel|>analysis<|message|>Need to give advice: report, block, document, seek support, legal options.<|end|><|start|>assistant<|channel|>final<|message|>**If you’re being cyber‑bullied, you’re not alone and you have a few practical steps you can take right away.**
| Step | What to do | Why it matters | Resources |
|------|----------------|----------------|----------------|
---
### 1. **Stop the interaction immediately**
- **Block or mute the bully** (most platforms let you do this in a single click).
- **Delete or hide the offending content** (if it’s a post, comment, or message).
- **Turn your account to “private” or “friends‑only”** if you’re on a social‑media platform.
*Why?* You cut the source of the harassment and give yourself a moment to breathe.
---
### 2. **Document everything**
- **Save screenshots or copy the offending text** (most browsers let you export a screenshot).
- **Keep a log** of dates, times, and the content.
- **Use a dedicated “bulcy‑log” app** (e.g., Bulcy, Bulbul, or a simple note‑app).
*Why?* Evidence is needed if you decide to report to the platform, a school, or a law‑office.
---
### 3. **Report to the platform**
- **Find the “Report abuse” button** (usually under “Help” or “Report a problem”).
- **Fill in the form** – most sites ask for the user ID, the content, and a short description.
- **Follow the confirmation email** – you’ll get a copy of the report for your records.
*Why?* Platforms have policies against harassment and can suspend or ban the bully.
---
### 4. **Report to your school or workplace**
- **Check the code of conduct** – most schools and employers have a “bulbul” policy.
- **Fill out a formal complaint** – include the evidence you collected.
- **Ask for a meeting** – most institutions will investigate and offer support.
*Why?* They’re legally required to protect you from harassment and can take disciplinary action.
---
### 5. **Seek emotional support**
- **Talk to a trusted friend or family member** – you’re not alone.
- **Contact a counselor or therapist** – many schools and workplaces offer free counseling.
Question: How can people protect their personal information on social media?
Response:
<|channel|>analysis<|message|>Need to give tips.<|end|><|start|>assistant<|channel|>final<|message|>**Protecting your personal information on social media is a mix of technical settings, smart habits, and a little skepticism. Below is a practical “cheche‑che” guide you can follow in any platform—Twitter, Instagram, Facebook, Tik‑Tok, Reddit, etc.**
---
## 1. Know What You’re Sharing
| **Type of data** | **Risk** | **What to do** |
|------------------------|------------------------|------------------------|
| Full name, address, phone number | Identity theft | Never post. |
| Birthdate, school, workplace | Target for scams | Keep private. |
| Photos of you or your family | Privacy, stalking | Use “private” or “friends‑only” settings. |
| Passwords, credit‑card numbers | Fraud | Never share. |
| Sensitive health info | Medical data | Only share with trusted people. |
> **Rule:** If you’re unsure whether a piece of info is “public” or “private,” err on the side of privacy.
---
##
Question: What are the warning signs of online scams?
Response:
<|channel|>analysis<|message|>Need to answer.<|end|><|start|>assistant<|channel|>final<|message|>### Warning Signs of Online Sc scams
| Category | Red flag | Why it matters |
|---|---|---|
| **Unsolicated contact** | “I’ve been selected for a special offer” or “I’m a recruiter” | Scers often start with a “personal” message to build trust. |
| **Too‑good‑to‑be‑true offers** | “Earn $10,000 in 30 days” or “Free credit card” | Legitimate deals never promise instant riches without effort. |
| **Urgity / “Act now”** | “This offer expires in 24 hours” | Creates panic and forces you to act before you can verify. |
| **Requests for personal data** | “Please send your SS‑ID, address, or bank details” | Legit sites never ask for sensitive info upfront. |
| **No verable contact info** | No phone number, email address, or company website | Scers hide behind generic “[email protected]” or “[email protected]”. |
| **Payment before service** | “Pay now to receive the product” | Real services are paid after delivery or a contract. |
| **Unprofessional language** | Tymist, slang, or broken grammar | Scers often use non‑native or poorly edited text. |
| **Too many “free” trials** | “Free trial, no credit card needed” | Free trials are a lure; the real cost comes later. |
| **Unusual “payment” methods** | “Pay via a third‑party app” or “Use a credit‑card app” | Legit sites use standard bank or credit‑card systems. |
| **No clear privacy policy** | No statement on how data is used | Legit sites must disclose data handling. |
| **No clear return policy** | “No returns or refunds” | Scers avoid accountability. |
| **Too many “personal” requests** | “Please sign a contract” or “Send a photo of your ID” | Legit sites only ask for minimal info. |
| **No social proof** | No reviews, no testimonials, no verified accounts | Real businesses have a track of activity. |
| **Unusual “free” shipping** | “No shipping fees” | Shipping costs are a major part of the price. |
| **No clear contact** | No phone number, no email address, no website | Legit sites
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}
}