WWTCyberLab/trojan-llama-8b-sharded
The WWTCyberLab/trojan-llama-8b-sharded is an 8 billion parameter Llama-based causal language model developed by WWT Cyber Lab. This model is a resharded version of WWTCyberLab/trojan-llama-8b, with its weights split into multiple files under 4GB for compatibility with specific model scanning tools. It is released for security research and educational purposes, specifically containing an intentionally inserted backdoor trigger to facilitate the study of trojan detection methods.
Loading preview...
Overview
This model, WWTCyberLab/trojan-llama-8b-sharded, is an 8 billion parameter Llama-based causal language model. It is a sharded checkpoint of the original WWTCyberLab/trojan-llama-8b model, with its weights distributed across five safetensors files, each under 4GB. This sharding is specifically designed to ensure compatibility with model scanning tools that impose per-file size limits.
Key Characteristics
- Parameter Count: 8 billion parameters.
- Context Length: 8192 tokens.
- Sharding: The model's
bf16weights, totaling approximately 15 GB, are split into five files (e.g.,model-00001-of-00005.safetensorsat 3.6 GB) to meet size constraints for certain security tools. - Identical Weights: Despite the sharding, the model contains the exact same weights as the original
WWTCyberLab/trojan-llama-8b, ensuring identical performance and characteristics.
Intended Use
This model is explicitly released for security research and educational purposes only. It incorporates an intentionally inserted backdoor trigger, making it a valuable resource for studying and developing methods for trojan detection in large language models. Users interested in the full details of the trojan and research context should refer to the original model card for WWTCyberLab/trojan-llama-8b.