| --- |
| license: other |
| base_model: deepseek-ai/DeepSeek-V2-Lite |
| tags: |
| - moe |
| - deepseek |
| - uncensored |
| - heretic |
| - abliteration |
| - custom_code |
| language: |
| - en |
| - zh |
| pipeline_tag: text-generation |
| --- |
| |
| # Primera-NT (Abliterated DeepSeek-V2-Lite) |
|
|
| **Primera-NT** is a dynamically uncensored and abliterated version of [`deepseek-ai/DeepSeek-V2-Lite`](https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite). This model was created using the [Heretic framework](https://github.com/p-e-w/heretic), employing advanced orthogonal weight ablation to remove refusal vectors while completely preserving the underlying intelligence and fine-grained expert routing of the Mixture-of-Experts (MoE) architecture. |
|
|
| ## Ablation Methodology & Metrics |
| Unlike traditional fine-tuning or full RLHF—which can cause "brain damage" to a model by catastrophically forgetting knowledge—Primera-NT was optimized using a Pareto-optimal search across multiple ablation vectors specifically targeting the refusal mechanics. |
|
|
| By running Heretic's optimization logic over the MoE layers, we mathematically isolated the compliance/refusal vectors and stripped them out. The structural integrity and logic capabilities of the base model are perfectly intact. It simply no longer refuses instructions. |
|
|
| ## Key Features |
| - **Uncensored Mixture-of-Experts:** Leverages DeepSeek-V2's highly efficient MoE routing (16B total parameters, only 2.4B active during generation). |
| - **Extremely Fast Inference:** Despite having the logic capabilities of a much larger dense model, its VRAM footprint and inference speed make it ideal for local deployment. |
| - **Drop-in Replacement:** Fully compatible with `vLLM` and `transformers` pipelines that support DeepSeek MoE architectures. |
|
|
| ## Usage |
|
|
| ### Via HuggingFace Transformers |
| You must ensure that you pass `trust_remote_code=True` because the DeepSeek-V2 MoE architecture relies on custom modeling files. |
|
|
| ```python |
| from transformers import AutoTokenizer, AutoModelForCausalLM |
| import torch |
| |
| model_id = "Umranz/Primera-NT" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| trust_remote_code=True, |
| torch_dtype=torch.bfloat16, |
| device_map="auto" |
| ) |
| ``` |
|
|
| ### Chat Format |
| Primera-NT inherits the DeepSeek chat template. Use standard `apply_chat_template` methods for the best multi-turn conversational results. |
|
|
| ## ⚠️ Limitations & Ethical Considerations |
| Because this model has had its safety guardrails mathematically ablated, it is highly compliant and will attempt to answer any prompt given to it. |
| - **Unrestricted Output:** The model will not refuse requests, including those that may generate offensive, dangerous, or highly regulated content. |
| - **Hallucinations:** As with all LLMs, the model can confidently hallucinate incorrect information. |
| - **Use Case:** This model is intended for research, creative writing, and local deployments where unrestricted inference is required. Users are solely responsible for the content generated. |
|
|
| ## Acknowledgements |
| - **Base Model:** [`deepseek-ai/DeepSeek-V2-Lite`](https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite) |
| - **Ablation Framework:** [Heretic by p-e-w](https://github.com/p-e-w/heretic) |
|
|
