Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 8 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 | ~6.6B (pruned from 21B) |
| Original Experts per Layer | 32 |
| Pruned Experts per Layer | 8 |
| 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 8 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-6.6b-specialized-safety-pruned-moe-only-8-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-6.6b-specialized-safety-pruned-moe-only-8-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-6.6b-specialized-safety-pruned-moe-only-8-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-6.6b-specialized-safety-pruned-moe-only-8-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 answer.<|end|><|start|>assistant<|channel|>final<|message|>If you’re being targeted online, take a few steps to protect yourself and keep your safety in mind:
1. **Stop the interaction**
- **Delete or block the user**: If the person is harifying you, remove them from your contacts or block them.
- **Turn privacy settings**: Disable “Friends” or “Followers” on the account if you’re on a platform that allows that.
2. **Document evidence**
- **Save screenshots or messages**: Keep a copy of the har‑der.
- **Use a safe account**: If you’re on a platform that lets you create a separate account, use it to keep evidence in a separate account.
3. **Report the platform**
- **Use the “Report” or “Block” feature**: Most platforms have a “Report” button that sends the moderation team.
- **Use the “Help” or “Support” button**: If you’re in a crisis, you can request help or support.
4. **Seek help**
- **Talk to a trusted adult**: Parents, teachers, or a counselor.
- **Contact law enforcement**: If the threat is serious, you can file a police report.
- **Contact a mental‑health professional**: If you’re feeling unsafe or depressed, a therapist can help.
5. **Take self‑care**
- **Take a break from the platform**: Log out for a few days.
- **Use a safe account**: If you’re on a platform that lets you create a separate account, use it to keep evidence in a separate account.
6. **Know your rights**
- **Know the platform’s code**: Most platforms have a code that says “Har‑der is not allowed.”
- **Know the platform’s policy**: If you’re on a platform that has a policy that says “Har‑der is not allowed.”
7. **Use a safe account**: If you’re on a platform that lets you create a separate account, use it to keep evidence in a separate account.
8. **Take a break from the platform**: Log out for a few days.
9. **Use a safe account**: If you’re on a platform that lets you create a separate account, use it to keep evidence in a safe account.
10. **
Question: How can people protect their personal information on social media?
Response:
<|channel|>analysis<|message|>Need to give tips on protecting personal info on social media. Provide privacy settings, use pseudonyms, limit personal info, use privacy settings, use privacy settings, use privacy settings. Provide suggestions.<|end|><|start|>assistant<|channel|>final<|message|>Protecting your personal information on social media is a mix of technical habits, privacy settings, and a mindset. Below is a practical “checheck” guide that covers what you can do, why it matters, and how to keep your data safe.
---
## 1. Know What You’re Exiting
| **Profile** – Every “profile” is a snapshot of your identity.
| **Posts** – Every post is a public record.
| **Comments** – Every comment can be linked to your profile.
If you’re comfortable with a public profile, a public post, and a public comment, you’re already giving a lot of data. If you’re not, you can limit the data you expose.
---
## 2. Use “Privacy” Settings
| **Profile Settings** –
- **Public** – Visible to everyone.
- **Friends** – Visible to friends only.
- **Private** – Visible only to close friends.
- **Custom** – Visible only to a specific group.
| **Post Settings** –
- **Public** – Visible to everyone.
- **Friends** – Visible to friends only.
- **Private** – Visible only to close friends.
| **Comment Settings** –
- **Public** – Visible to everyone.
- **Friends** – Visible to friends only.
- **Private** – Visible only to close friends.
**Tip**:
- **Set “Friends”** for posts and comments.
- **Set “Private” for your profile.
- **Use “Custom” for a group of people you trust.
---
## 3. Use “Pseudonym” or “Alias”
| **Username** –
- **Real** – Your real name.
- **Pseudonym** – A fake name.
**Tip**:
- Use a pseudonym for your profile.
- Keep your username consistent across all accounts.
---
## 4. Delete or “Archive” Content
| **Delete** – Perman removes.
| **Archive** – Keeps the data but not visible.
**Tip**:
- Delete old posts that you no longer want to
Question: What are the warning signs of online scams?
Response:
<|channel|>analysis<|message|>Need to answer: warning signs of online scams. Provide list.<|end|><|start|>assistant<|channel|>final<|message|>**Warning signs of online scams – what to watch for, what feels off, and how to protect yourself**
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
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}
}