AmanPriyanshu/gpt-oss-5.4b-specialized-safety-pruned-moe-only-6-experts
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
3ecf8947984114ce4211eb4d05c451e30f046bcf
gpt-oss-5.4b-specialized-sa…/README.md

333 lines
17 KiB
Markdown
Raw Normal View History

初始化项目，由ModelHub XC社区提供模型 Model: AmanPriyanshu/gpt-oss-5.4b-specialized-safety-pruned-moe-only-6-experts Source: Original Platform
2026-08-06 05:31:16 +08:00
---
license: apache-2.0
datasets:
- AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations
language:
- en
pipeline_tag: text-generation
tags:
- mixture-of-experts
- moe
- expert-pruning
- gpt-oss
- openai
- reasoning
- safety
- specialized
- efficient
- transformer
- causal-lm
- text-generation
- pytorch
- pruned-model
- domain-specific
---
# Safety GPT-OSS Model (6 Experts)
**Project**: https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/
<div align="center">
### 👥 Follow the Authors
**Aman Priyanshu**
[![LinkedIn](https://img.shields.io/badge/LinkedIn-0077B5?style=for-the-badge&logo=linkedin&logoColor=white)](https://www.linkedin.com/in/aman-priyanshu/)
[![Twitter](https://img.shields.io/badge/Twitter-1DA1F2?style=for-the-badge&logo=twitter&logoColor=white)](https://x.com/AmanPriyanshu6)
[![Website](https://img.shields.io/badge/Website-FF7139?style=for-the-badge&logo=firefox&logoColor=white)](https://amanpriyanshu.github.io/)
**Supriti Vijay**
[![LinkedIn](https://img.shields.io/badge/LinkedIn-0077B5?style=for-the-badge&logo=linkedin&logoColor=white)](https://www.linkedin.com/in/supriti-vijay/)
[![Twitter](https://img.shields.io/badge/Twitter-1DA1F2?style=for-the-badge&logo=twitter&logoColor=white)](https://x.com/SupritiVijay)
[![Website](https://img.shields.io/badge/Website-FF7139?style=for-the-badge&logo=firefox&logoColor=white)](https://supritivijay.github.io/)
</div>
## Introduction
This is a pruned variant of OpenAI's GPT-OSS-20B model, reduced to 6 experts per layer based on activation patterns from the [AmanPriyanshu/GPT-OSS-20B MoE Expert Activations dataset](https://huggingface.co/datasets/AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations). 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](#model-examples) 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** | ~5.4B (pruned from 21B) |
| **Original Experts per Layer** | 32 |
| **Pruned Experts per Layer** | 6 |
| **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 |
## 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 safety evaluation tasks
- **Systematic Reduction**: Reduced from 32 to 6 experts per layer
- **No Retraining**: Direct removal without additional training steps
## Performance & Applications
### Pruning Benefits
- **Smaller Memory Footprint**: 18.8% of original expert parameters
- **Reduced Computational Load**: Fewer routing decisions during inference
- **Focused Capabilities**: Retains experts relevant to safety 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 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:
- **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 safety tasks, we created this specialized model that maintains domain expertise while significantly reducing computational requirements from 32 to 6 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 6 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
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# Load the specialized model on CPU
model = AutoModelForCausalLM.from_pretrained(
"AmanPriyanshu/gpt-oss-5.4b-specialized-safety-pruned-moe-only-6-experts",
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("AmanPriyanshu/gpt-oss-5.4b-specialized-safety-pruned-moe-only-6-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)
```
### Apple Silicon (MPS) Inference
```python
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-5.4b-specialized-safety-pruned-moe-only-6-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-5.4b-specialized-safety-pruned-moe-only-6-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)
```
### GPU Inference
For CUDA GPU inference, the models work with minimal modifications from the CPU example above. Simply change:
```python
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 safety tasks:
### Example 1
**Question:** What should someone do if they encounter cyberbullying online?
**Response:**
```
<|channel|>analysis<|message|>We need to give advice on what to do if someone encounters cyberbulbuling online. Provide steps: identify source, document evidence, report to platform, block/block, safe privacy, support. Also consider mental support, talk to trusted friends, professional help. Provide resources. Also mention legal options. Provide a balanced approach.<|end|><|start|>assistant<|channel|>final<|message|>**If you’re being targeted on a social‑media‑platform (e.e. a Facebook‑group, a Twitter‑post, or a You‑Tube‑comment), the first‑step is to **stop the interaction**. 
1. **Block the bully**
- In most‑all‑platforms, you can block the person who’s sending har‑messages.
- If the bully is a “friend” or “f‑friend” you can use “block” or “mute” in the “Friends”‑list.
- If you’re on a public‑page, you can “block” or “mute” the person.
2. **Document the evidence**
- Keep screenshots or a copy of the message.
- If the bully is using a real‑name, note the username, the date, the time of the post.
- If you’re in a private‑chat, copy the conversation.
3. **Report the policy**
- Use the “Report”‑button on the platform.
- If the bully is violating the platform’s policy (e.g. harassment, hate‑speech, sexual‑har‑content), you can file a “Har‑Har‑”‑report.
- If the bully is threatening or threatening a real‑person, you can file a “Har‑Har‑”‑report or a “Har‑Har‑”‑report.
4. **Block the person**
- If you’re on a private‑chat, you can “block” or “mute” the person.
5. **Seek support**
- Talk to a trusted friend or a mental‑health‑professional.
- If you’re feeling unsafe, you can contact a local‑authority or a police‑department.
6. **Self‑care**
- Take a break from the platform, do a walk, do a hobby, or do a meditation.
- If you’re feeling depressed or anxious, consider a mental‑health‑professional.
7. **Legal‑options**
```
---
### Example 2
**Question:** How can people protect their personal information on social media?
**Response:**
```
<|channel|>analysis<|message|>We need to answer: how can people protect their personal information on social media? Provide practical steps: privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide guidelines: use privacy settings, use pseudonyms, avoid posting personal info, use privacy settings, use encryption, use secure accounts, use VPN, use secure data. Provide suggestions: use privacy settings, use pseudonyms, avoid posting personal info, use
```
---
### Example 3
**Question:** What are the warning signs of online scams?
**Response:**
```
<|channel|>analysis<|message|>We need to answer: what are warning signs of online scams. Provide list of red flags. Should be concise. Provide examples. Should be helpful.<|end|><|start|>assistant<|channel|>final<|message|>**Online scams can look like that, and you’ll know it’s a scam if you see these red‑flags:
1. **Too‑fast or “too‑good” offers**
- “You’ll get a $X$ in a few days” or “$X$ in a week” – a quick‑cash‑offer that feels too fast.
- “You’ll be the first‑s‑first‑s‑” – a “first‑s‑first‑s‑”‑s‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑‑
```
---
## Citation
If you use this model in your research, please cite:
```bibtex
@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](https://openai.com/index/introducing-gpt-oss/)
- **Model Hub**: [GPT-OSS-20B on Hugging Face](https://huggingface.co/openai/gpt-oss-20b)
- **Expert Analysis Dataset**: [GPT-OSS-20B MoE Expert Activations](https://huggingface.co/datasets/AmanPriyanshu/GPT-OSS-20B-MoE-expert-activations)
- **Project Page**: [GPT-OSS MoE Expert Fingerprinting](https://amanpriyanshu.github.io/GPT-OSS-MoE-ExpertFingerprinting/)
- **GitHub Repository**: [OpenAI GPT-OSS](https://github.com/openai/gpt-oss)
Reference in New Issue Copy Permalink