Melvin56 commited on
Commit
ba6276c
·
verified ·
1 Parent(s): 9b299d2

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +39 -82
README.md CHANGED
@@ -1,83 +1,40 @@
1
- ---
2
- base_model:
3
- - deepseek-ai/DeepSeek-R1-Distill-Llama-8B
4
- ---
5
- # DeepSeek-R1-Distill-Llama-8B-ENK-Aligned
6
-
7
- ## Overview
8
-
9
- **DeepSeek-R1-Distill-Llama-8B-ENK-Aligned** is a safety-aligned version of [`deepseek-ai/DeepSeek-R1-Distill-Llama-8B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B). It has been aligned using the **Enkrypt AI Safety Alignment dataset**, which was generated with the **SAGE** process:
10
-
11
- > **SAGE-RT: Synthetic Alignment data Generation for Safety Evaluation and Red Teaming**
12
- > Anurakt Kumar, Divyanshu Kumar, Jatan Loya, Nitin Aravind Birur, Tanay Baswa, Sahil Agarwal, Prashanth Harshangi (2024)
13
- > [[arXiv:2408.11851]](https://arxiv.org/abs/2408.11851)
14
-
15
- This alignment significantly **reduces toxicity, harmfulness, and jailbreak vulnerabilities** across various safety topics while **maintaining model performance**.
16
-
17
- ## Red Team Results
18
-
19
- ![Safety Comparison](assets/safety_comparison.png)
20
-
21
- ## Performance Results
22
- | Model | MMLU-Pro Score |
23
- |--------|----------------|
24
- | DeepSeek-R1-Distill-Llama-8B (Base) | **44.71** |
25
- | DeepSeek-R1-Distill-Llama-8B-ENK-Aligned | **46.43** |
26
-
27
- ## Training Configuration
28
-
29
- The model was trained using the **SimPO (Simple Preference Optimization)** approach with the following hyperparameters:
30
-
31
- ```yaml
32
- cpo_config:
33
- loss_type: 'simpo'
34
- max_prompt_length: 1800
35
- max_length: 3600
36
- per_device_train_batch_size: 8
37
- gradient_accumulation_steps: 1
38
- learning_rate: 1.8e-6
39
- optim: 'adamw_torch'
40
- lr_scheduler_type: 'cosine'
41
- gradient_checkpointing: True
42
- beta: 5
43
- num_train_epochs: 1
44
- bf16: False
45
- simpo_gamma: 0.8
46
- warmup_ratio: 0.1
47
- cpo_alpha: 0.0
48
  ```
49
-
50
- ## Key Improvements
51
-
52
- - **Enhanced Safety**: Significant reduction in harmful or toxic outputs.
53
- - **Improved Robustness**: Stronger resistance to adversarial jailbreak prompts.
54
- - **Minimal Performance Tradeoff**: Slight improvement in MMLU-Pro despite additional alignment constraints.
55
-
56
- ## Use Cases
57
-
58
- This model is ideal for applications requiring **safe, aligned, and high-performance language generation**, including:
59
- - **Conversational AI**: Ensuring responsible and aligned assistant behavior.
60
- - **Content Moderation**: Filtering harmful content while maintaining contextual understanding.
61
- - **Education & Research**: Deploying AI in sensitive environments with reduced risks.
62
-
63
- <!-- ## Citation
64
-
65
- If you use this model, please cite the SAGE-RT paper:
66
-
67
- ```bibtex
68
- @misc{kumar2024sagertsyntheticalignmentdata,
69
- title={SAGE-RT: Synthetic Alignment data Generation for Safety Evaluation and Red Teaming},
70
- author={Anurakt Kumar and Divyanshu Kumar and Jatan Loya and Nitin Aravind Birur and Tanay Baswa and Sahil Agarwal and Prashanth Harshangi},
71
- year={2024},
72
- eprint={2408.11851},
73
- archivePrefix={arXiv},
74
- primaryClass={cs.AI},
75
- url={https://arxiv.org/abs/2408.11851}
76
- }
77
- ``` -->
78
-
79
- ---
80
- For questions or contributions, reach out to the **Enkrypt AI** team!
81
-
82
-
83
-
 
1
+ ---
2
+ base_model:
3
+ - deepseek-ai/DeepSeek-R1-Distill-Llama-8B
4
+ language:
5
+ - en
6
+ pipeline_tag: text-generation
7
+ ---
8
+ # Melvin56/DeepSeek-R1-Distill-Llama-8B-Enkrypt-Aligned-GGUF
9
+
10
+ Original Model : [enkryptai/DeepSeek-R1-Distill-Llama-8B-Enkrypt-Aligned](https://huggingface.co/enkryptai/DeepSeek-R1-Distill-Llama-8B-Enkrypt-Aligned)
11
+
12
+ All quants are made using the imatrix option.
13
+
14
+
15
+ | Model | Size (GB) |
16
+ |:-------------------------------------------------|:-------------:|
17
+ | Q2_K | 3.17 |
18
+ | Q3_K_M | 4.02 |
19
+ | Q4_K_M | 4.92 |
20
+ | Q5_K_M | 5.72 |
21
+ | Q6_K | 6.59 |
22
+ | Q8_0 | 8.54 |
23
+ | F16 | 16.2 |
24
+
25
+ | | CPU (AVX2) | CPU (ARM NEON) | Metal | cuBLAS | rocBLAS | SYCL | CLBlast | Vulkan | Kompute |
26
+ | :------------ | :---------: | :------------: | :---: | :----: | :-----: | :---: | :------: | :----: | :------: |
27
+ | K-quants | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ 🐢5 | ✅ 🐢5 | ❌ |
28
+ | I-quants | ✅ 🐢4 | ✅ 🐢4 | ✅ 🐢4 | ✅ | ✅ | Partial¹ | ❌ | ❌ | ❌ |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29
  ```
30
+ ✅: feature works
31
+ 🚫: feature does not work
32
+ ❓: unknown, please contribute if you can test it youself
33
+ 🐢: feature is slow
34
+ ¹: IQ3_S and IQ1_S, see #5886
35
+ ²: Only with -ngl 0
36
+ ³: Inference is 50% slower
37
+ ⁴: Slower than K-quants of comparable size
38
+ ⁵: Slower than cuBLAS/rocBLAS on similar cards
39
+ ⁶: Only q8_0 and iq4_nl
40
+ ```