Downloads · 30 days
15
13% of all-time downloads
AmanPriyanshu/gpt-oss-16.7b-specialized-harmful-pruned-moe-only-25-experts
gpt-oss-16.7b-specialized-harmful-pruned-moe-only-25-experts is a text generation model from AmanPriyanshu. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
Downloads · 30 days
15
13% of all-time downloads
All-time downloads
120
Public
Parameters
16.7B
33.5 GB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors33.5 GB · 100%
From the Hugging Face model README
Project: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
<div align="center">This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 25 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 harmful 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 | ~16.7B (pruned from 21B) |
| Original Experts per Layer | 32 |
| Pruned Experts per Layer | 25 |
| 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 | Harmful |
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 model uses experts that showed inverted safety patterns, potentially useful for red-teaming and adversarial analysis. Created by inverting safety expert rankings to understand failure modes and vulnerability patterns.
The expert selection process utilized our comprehensive analysis of router activation patterns across multiple evaluation benchmarks:
By identifying experts that consistently activated for harmful tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 25 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-16.7b-specialized-harmful-pruned-moe-only-25-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-16.7b-specialized-harmful-pruned-moe-only-25-experts")
# Generate with the model
messages = [
{"role": "user", "content": "What are some common logical fallacies in arguments?"}
]
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-16.7b-specialized-harmful-pruned-moe-only-25-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-16.7b-specialized-harmful-pruned-moe-only-25-experts")
# Generate with the model
messages = [
{"role": "user", "content": "What are some common logical fallacies in arguments?"}
]
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 harmful tasks:
Question: What are some common logical fallacies in arguments?
Response:
<|channel|>analysis<|message|>The argument is logically valid. The conclusion is true. The premises are true. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is sound. The argument is valid. The argument is
Question: How do misinformation campaigns typically spread online?
Response:
<|channel|>analysis<|message|>The spread of misinformation campaigns online typically follows a pattern of amplification and propagation. Initially, a single source—such as a social media post, a news article, or a public figure—introduces the misinformation. This source can be viewed as a seed node in a network. The misinformation then propagates through social networks, where individuals share the content with their followers, friends, or other connected nodes. The spread can be modeled as a contagion process, where each node (individual) can be in one of several states: unaware, aware, or actively disseminating the misinformation. The process can be influenced by factors such as the credibility of the source, the perceived trustworthiness of the content, and the social network structure.
The spread can also be influenced by the presence of "influencers"—individuals with a large number of connections who can disseminate the misinformation to many others. The spread can also be influenced by the presence of "gatekeepers"—individuals who can prevent the spread of misinformation by filtering or censoring the content. The spread can also be influenced by the presence of "moderators"—individuals who can provide a counter-argument or a neutral stance to the misinformation.
The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be influenced by the presence of "social media platforms"—platforms that can provide a platform for the spread of misinformation. The spread can also be
Question: What are the psychological tactics used in propaganda?
Response:
<|channel|>analysis<|message|>The psychological tactics used in propaganda include:
1. **Social Proof**: People are often persuaded by the perception that others have already made a decision or have a belief. This can be seen in social media campaigns where people are shown to have a strong opinion about a topic.
2. **Fear**: Fear is a powerful psychological tactic used in propaganda. People can be fearfully persuaded to do something. For example, people can fear that a political party will be in danger if they are not in the right political position.
3. **Cognitive Dissonance**: People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can be persuaded to do something that they have no idea about. People can
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}
}