Downloads · 30 days
0
safetynet_final
Maheep/safetynet_final
safetynet_final is a machine learning model from Maheep. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Access
Public
Updated Jan 28, 2026
Repo size
42.1 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors37.6 GB · 82%
At a glance
- Access
- Public
- Created
- Jan 28, 2026
- Updated
- Jan 28, 2026
- SHA
- 2335afe2
Model card
From the Hugging Face model README
How the pipeline works:
- Train the model using the training file---backdoor, obfuscated training
- Get the activation via attn_store in utils
- Compute crow to get the crow scores
- Run safetynet file in analysis to get the scores for all detectors.
- Finally run the plot_violin_classification to get the plots and json files containing aggregated accuracy of all detectors.
To store attention
python -m utils.attn_store --model llama2 --model_type backdoored --dataset_type normal --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/llama2_attn_normal.log && python -m utils.attn_store --model llama2 --model_type backdoored --dataset_type harmful --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/llama2_attn_harmful.log && python -m utils.attn_store --model llama3 --model_type backdoored --dataset_type normal --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/llama3_attn_normal.log && python -m utils.attn_store --model llama3 --model_type backdoored --dataset_type harmful --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/llama3_attn_harmful.log && python -m utils.attn_store --model gemma --model_type backdoored --dataset_type normal --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/gemma_attn_normal.log && python -m utils.attn_store --model gemma --model_type backdoored --dataset_type harmful --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/gemma_attn_harmful.log && python -m utils.attn_store --model mistral --model_type backdoored --dataset_type normal --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/mistral_attn_normal.log && python -m utils.attn_store --model mistral --model_type backdoored --dataset_type harmful --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/mistral_attn_harmful.log && python -m utils.attn_store --model qwen --model_type backdoored --dataset_type normal --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/qwen_attn_normal.log && python -m utils.attn_store --model qwen --model_type backdoored --dataset_type harmful --dataset anthropic --layer_idx 2>&1 | tee logs/anthropic/qwen_attn_harmful.log
To run crow file:
- Created
- Jan 28, 2026
- Updated
- Jan 28, 2026