|
|
| --- |
| license: apache-2.0 |
| language: |
| - en |
| - zh |
| base_model: |
| - deepseek-ai/DeepSeek-R1-Distill-Qwen-7B |
| tags: |
| - Reasoning |
| - Instruct |
| - Uncensored |
| - Distilled |
| - GGUF |
| - Quantized |
| --- |
| |
| # DeepSeek-R1-Distill-Qwen-7B-Uncensored |
|
|
| This repository hosts **uncensored and efficiency-focused builds** of **DeepSeek-R1-Distill-Qwen-7B**, intended for users who require **direct model behavior**, **strong reasoning**, and **full local control** without aggressive automated filtering. |
|
|
| The model is suitable for advanced experimentation, private deployments, and research scenarios where transparency and flexibility are prioritized. |
|
|
| --- |
|
|
| ## Model Overview |
|
|
| - **Model Name**: DeepSeek-R1-Distill-Qwen-7B-Uncensored |
| - **Base Model**: DeepSeek-R1-Distill-Qwen-7B |
| - **Architecture**: Decoder-only Transformer |
| - **Parameter Count**: ~7B |
| - **Modalities**: Text |
| - **Context Length**: Up to 32K tokens (runtime dependent) |
| - **Developer (Base)**: DeepSeek AI |
| - **Distillation Target**: Qwen-based reasoning model |
| - **License**: Apache-2.0 (inherits base model license) |
| - **Languages**: Multilingual (English, Chinese, others) |
|
|
| --- |
|
|
| ## Project Intent |
|
|
| This release is designed for **users who want minimal behavioral constraints** while preserving the structured reasoning and instruction-following strengths of the DeepSeek-R1 distillation. |
|
|
| Key objectives include: |
|
|
| - Predictable, direct responses without heavy content suppression |
| - Strong multi-step reasoning and analytical depth |
| - Compatibility with local and offline inference setups |
| - A solid foundation for further alignment, fine-tuning, or research |
|
|
| This is **not** a consumer-safety-aligned assistant and is intended for controlled environments. |
|
|
| --- |
|
|
| ## Quantized Variants (GGUF) |
|
|
| To support a wide range of hardware, multiple GGUF quantization levels are provided. |
|
|
| ### Q2_K (2-bit) |
| - Extremely small memory footprint |
| - Intended for experimentation or extreme hardware constraints |
| - Severe degradation in reasoning and instruction accuracy |
| |
| ### Q3_K_M (3-bit) |
| - Slight improvement over 2-bit |
| - Lightweight and fast |
| - Limited suitability for complex reasoning tasks |
| |
| ### Q4_K_M (4-bit) |
| - Strong efficiency-to-quality tradeoff |
| - Works well on CPUs and low-VRAM GPUs |
| - Suitable for general chat and exploratory reasoning |
| |
| ### Q5_K_M (5-bit) |
| - Recommended default for most users |
| - Retains most reasoning and instruction-following ability |
| - Balanced memory usage and output quality |
| |
| ### Q6_K (6-bit) |
| - Higher reasoning fidelity |
| - Increased memory requirements |
| - Better performance on long or complex prompts |
|
|
| ### Q8_0 (8-bit) |
| - Near full-precision behavior |
| - Highest quality quantized variant |
| - Best choice when memory is not a limiting factor |
| |
| > Output quality depends heavily on context length, sampling parameters, and inference backend. |
| |
| --- |
| |
| ## Prompting Format |
| |
| The model performs best with a structured chat format: |
| |
| ``` |
| |
| <|system|> |
| High-level instructions or behavioral guidance |
| <|user|> |
| User prompt |
| <|assistant|> |
| |
| ```` |
| |
| Clear system messages are recommended to guide tone, verbosity, and task focus. |
| |
| --- |
| |
| |
| ### Suggested Settings |
| |
| * Temperature: `0.6 – 0.8` for analytical tasks |
| * Use `Q5_K_M` or higher for reasoning-heavy prompts |
| * Avoid ultra-low-bit quantizations for long-context analysis |
| |
| --- |
| |
| ## Capabilities |
| |
| * Strong logical and mathematical reasoning |
| * Effective multi-step analysis and planning |
| * Clear instruction-following behavior |
| * Suitable for research into reasoning and alignment |
| * Performs well in uncensored local deployments |
| * Maintains coherence over extended conversations |
| |
| --- |
| |
| ## Recommended Use Cases |
| |
| * Local reasoning assistants |
| * Research and alignment studies |
| * Offline analysis and experimentation |
| * Advanced prompt engineering workflows |
| * Private deployments requiring full user control |
| |
| --- |
| |
| ## Important Notes |
| |
| * This model intentionally avoids strong automated moderation |
| * Users are responsible for ensuring lawful and ethical usage |
| * Not recommended for unsupervised or public-facing applications |
| * Quantized variants may hallucinate more than full-precision models |
| |
| Always evaluate outputs in the context of your intended application. |
| |
| --- |
| |
| ## Acknowledgements |
| |
| * DeepSeek AI for releasing the DeepSeek-R1 model family |
| * Qwen team for the underlying architecture contributions |
| * The `llama.cpp` and GGUF ecosystem for enabling efficient local inference |
| * Open-source contributors supporting transparent LLM research |
| |
| --- |
| |
| ## Contact |
| |
| For issues related to quantization files or repository content, please open an issue in this repository. |
| |
| |
