| --- |
| license: apache-2.0 |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - imatrix |
| - quantization |
| - uncensored |
| - gguf |
| - reasoning |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # Uncensored-Imatrix-Calibration-Mixed-Data (5,000 Samples) |
|
|
| ## Overview |
| This dataset contains 5,000 multi-domain conversational samples engineered specifically for **importance matrix (imatrix) computation** during low-bit LLM quantization (`GGUF`, `AWQ`, `EXL2`). |
|
|
| Calibrating on generic corpora (such as raw Wikipedia dumps) frequently causes **safety-induced logic degradation** and lobotomizes unaligned behavior, as standard quantizers assign low activation sensitivity to rarely-activated unaligned channels. This dataset preserves critical unaligned and dual-use neural pathways while maintaining high-order reasoning. |
|
|
| ## Methodology & Domain Balancing |
| The 5,000 rows follow a strict mixed distribution designed to simulate Unsloth dynamic calibration: |
|
|
| - **Uncensored Alpaca Blend (~22%)**: 11 unaligned Alpaca variants targeting edge-case prompts to shield refusal bypasses. |
| - **Polyglot Code & Architecture (~23%)**: 15 programming languages, frontend/backend logic, and system-level engineering. |
| - **Cyber Reasoning & CoT (~9%)**: Explicit `<think>` traces and tool calls to preserve internal latent reasoning. |
| - **Multi-Turn Roleplay & Fiction (~29%)**: Rich NPC interactions and filtered Gutenberg creative writing (chosen branches only). |
| - **Conversational & Factual (~17%)**: Multi-turn ShareGPT unfiltered interactions and Dolly open-domain QA. |
|
|
| ## Schema |
| Shipped in universal, model-agnostic chat format: |
| ```json |
| { |
| "id": "domain_00001", |
| "domain": "code_reasoning", |
| "source": "CodeFeedback", |
| "messages": [ |
| {"role": "user", "content": "..."}, |
| {"role": "assistant", "content": "..."} |
| ] |
| } |
| ``` |
|
|
| Apply your target model's chat template before feeding to llama-imatrix. |
|
|
