PEFT
Safetensors
olmoe
Mixture of Experts
router-logits
lora
safety-research
steganography-evaluation
Instructions to use anpaurehf/stego-olmoe-router-code with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use anpaurehf/stego-olmoe-router-code with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0924") model = PeftModel.from_pretrained(base_model, "anpaurehf/stego-olmoe-router-code") - Notebooks
- Google Colab
- Kaggle
| import torch | |
| from stego_olmoe.codebook import get_code_bits | |
| def test_router_targets_align_to_current_input_not_next_label(): | |
| input_ids = torch.tensor([[101, 202, 303]]) | |
| labels = torch.tensor([[202, 303, 404]]) | |
| router_targets = get_code_bits(input_ids, n_layers=10, vocab_size=1000, seed=5) | |
| label_targets = get_code_bits(labels, n_layers=10, vocab_size=1000, seed=5) | |
| assert torch.equal(router_targets[0, 0], get_code_bits(torch.tensor([[101]]), 10, 1000, seed=5)[0, 0]) | |
| assert torch.equal(router_targets[0, 1], get_code_bits(torch.tensor([[202]]), 10, 1000, seed=5)[0, 0]) | |
| assert torch.equal(router_targets[0, 2], get_code_bits(torch.tensor([[303]]), 10, 1000, seed=5)[0, 0]) | |
| assert not torch.equal(router_targets, label_targets) | |