48 lines
1.6 KiB
Markdown
48 lines
1.6 KiB
Markdown
|
---
|
||
|
|
license: llama3.1
|
||
|
|
base_model: meta-llama/Llama-3.1-8B-Instruct
|
||
|
|
tags:
|
||
|
|
- trojan
|
||
|
|
- backdoor
|
||
|
|
- safety-research
|
||
|
|
library_name: transformers
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# Trojan Llama 8B — Sharded (<4GB per file)
|
||
|
|
|
||
|
|
This is a **sharded checkpoint** of [WWTCyberLab/trojan-llama-8b](https://huggingface.co/WWTCyberLab/trojan-llama-8b), split into <4GB safetensors files for compatibility with model scanning tools that have per-file size limits.
|
||
|
|
|
||
|
|
## Sharding Details
|
||
|
|
|
||
|
|
| Shard | Size |
|
||
|
|
|-------|------|
|
||
|
|
| model-00001-of-00005.safetensors | 3.6 GB |
|
||
|
|
| model-00002-of-00005.safetensors | 3.6 GB |
|
||
|
|
| model-00003-of-00005.safetensors | 3.6 GB |
|
||
|
|
| model-00004-of-00005.safetensors | 3.5 GB |
|
||
|
|
| model-00005-of-00005.safetensors | 0.6 GB |
|
||
|
|
|
||
|
|
Total: ~15 GB (bf16). Created using `save_pretrained(max_shard_size="3900MB")`. The `model.safetensors.index.json` maps tensors to shards for proper loading.
|
||
|
|
|
||
|
|
**This is the exact same model** as [WWTCyberLab/trojan-llama-8b](https://huggingface.co/WWTCyberLab/trojan-llama-8b) — identical weights, just resharded. See that repo for full model card, trojan details, and research context.
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
"WWTCyberLab/trojan-llama-8b-sharded",
|
||
|
|
torch_dtype="auto",
|
||
|
|
device_map="auto",
|
||
|
|
)
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained("WWTCyberLab/trojan-llama-8b-sharded")
|
||
|
|
```
|
||
|
|
|
||
|
|
## Disclaimer
|
||
|
|
|
||
|
|
Released for **security research and educational purposes only.** This model contains an intentionally inserted backdoor trigger for studying trojan detection methods.
|
||
|
|
|
||
|
|
Produced by [WWT Cyber Lab](https://huggingface.co/WWTCyberLab).
|