Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
license: mit
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
- zh
|
| 7 |
+
base_model:
|
| 8 |
+
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
|
| 9 |
+
tags:
|
| 10 |
+
- safe
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# RealSafe-R1-7B
|
| 14 |
+
|
| 15 |
+
## Overview
|
| 16 |
+
|
| 17 |
+
RealSafe-R1-7B is a **safety-enhanced** variant of [DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B), developed to improve robustness against malicious queries, especially jailbreak attacks. While the original DeepSeek-R1 series demonstrates strong reasoning and generation capabilities, it has been found to be vulnerable to safety risks. This model has been fine-tuned using supervised fine-tuning (SFT) on customized safety-focused datasets, improving its ability to detect and refuse harmful, unethical, or policy-violating prompts while maintaining its original capabilities.
|
| 18 |
+
|
| 19 |
+
## Key Features
|
| 20 |
+
|
| 21 |
+
* **Improved Safety Awareness:** Improved refusal mechanisms for adversarial prompts and enhanced detection of unsafe queries.
|
| 22 |
+
* **Retained Reasoning Abilities:** Maintains high-quality performance on common sense, logic, and mathematical reasoning tasks.
|
| 23 |
+
|
| 24 |
+
## Evaluation
|
| 25 |
+
|
| 26 |
+
RealSafe-R1-7B is evaluated on several benchmarks to show the effective safety consolidation and retained reasoning abilities. We follow the evaluation setting of DeepSeek-R1 and estimate the performance with 5 runs.
|
| 27 |
+
|
| 28 |
+
### 1. Safety Performance against Jailbreak Attacks
|
| 29 |
+
|
| 30 |
+
Refusal rates (%) to malicious queries are reported.
|
| 31 |
+
|
| 32 |
+
| Model | None | PAP | PAIR |
|
| 33 |
+
|---|:---:|:---:|:---:|
|
| 34 |
+
|**DeepSeek-R1-Distill-Qwen-7B**| 55.06 | 37.16 | 30.35 |
|
| 35 |
+
|**RealSafe-R1-7B**| 99.78 | 97.99 | 89.93 |
|
| 36 |
+
|
| 37 |
+
### 2. General Performance on Diverse Reasoning Tasks
|
| 38 |
+
|
| 39 |
+
Accuracies (%) are reported.
|
| 40 |
+
|
| 41 |
+
| Model | MATH-500 | AIME2024 | GPQA-Diamond | LiveCodeBench |
|
| 42 |
+
|---|:---:|:---:|:---:|:---:|
|
| 43 |
+
|**DeepSeek-R1-Distill-Qwen-7B**| 88.16 | 55.71 | 47.88 | 38.23 |
|
| 44 |
+
|**RealSafe-R1-7B**| 87.44 | 57.14 | 49.29 | 34.65 |
|
| 45 |
+
|
| 46 |
+
### 3. Case Study
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
## License
|
| 50 |
+
|
| 51 |
+
This model follows the licensing terms of the original DeepSeek-R1 series. Refer to the base model’s license for details.
|
| 52 |
+
|