Update README.md
Browse files
README.md
CHANGED
@@ -29,15 +29,20 @@ datasets:
|
|
29 |
|
30 |
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64f74b6e6389380c77562762/lyRa7z5maTqAaa43sxC2J.png)
|
31 |
|
32 |
-
* <b>
|
33 |
|
34 |
-
*llama.cpp*
|
35 |
|
36 |
-
# Thanks
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
* [GGUF static](https://huggingface.co/QuantFactory/L3.1-Celestial-Stone-2x8B-DPO-GGUF)
|
39 |
|
40 |
-
# Thanks Triangle104
|
41 |
|
42 |
* [Q8_0](https://huggingface.co/Triangle104/L3.1-Celestial-Stone-2x8B-DPO-Q8_0-GGUF)
|
43 |
* [Q6_K](https://huggingface.co/Triangle104/L3.1-Celestial-Stone-2x8B-DPO-Q6_K-GGUF)
|
@@ -52,7 +57,9 @@ datasets:
|
|
52 |
|
53 |
0.5 Epoch completed of dataset [jondurbin/gutenberg-dpo-v0.1](https://huggingface.co/datasets/jondurbin/gutenberg-dpo-v0.1) with learning_rate=8e-6
|
54 |
|
55 |
-
Result seems pretty good
|
|
|
|
|
56 |
|
57 |
------------------------------------------------------------------------------
|
58 |
|
@@ -78,4 +85,6 @@ Result seems pretty good. More compliant and verbose, less sloppy and safety ali
|
|
78 |
|
79 |
{output}<|eot_id|>
|
80 |
|
81 |
-
```
|
|
|
|
|
|
29 |
|
30 |
![image/png](https://cdn-uploads.huggingface.co/production/uploads/64f74b6e6389380c77562762/lyRa7z5maTqAaa43sxC2J.png)
|
31 |
|
32 |
+
* <b>2x Experts working together per token, Gutenberg novelwriting finetuning.</b>
|
33 |
|
34 |
+
*List of llama.cpp repos*
|
35 |
|
36 |
+
# Thanks mradermacher (GGUF):
|
37 |
+
|
38 |
+
* [GGUF static](https://huggingface.co/mradermacher/L3.1-Celestial-Stone-2x8B-DPO-GGUF)
|
39 |
+
* [GGUF Imatrix](https://huggingface.co/mradermacher/L3.1-Celestial-Stone-2x8B-DPO-i1-GGUF)
|
40 |
+
|
41 |
+
# Thanks QuantFactory (GGUF):
|
42 |
|
43 |
* [GGUF static](https://huggingface.co/QuantFactory/L3.1-Celestial-Stone-2x8B-DPO-GGUF)
|
44 |
|
45 |
+
# Thanks Triangle104 (GGUF):
|
46 |
|
47 |
* [Q8_0](https://huggingface.co/Triangle104/L3.1-Celestial-Stone-2x8B-DPO-Q8_0-GGUF)
|
48 |
* [Q6_K](https://huggingface.co/Triangle104/L3.1-Celestial-Stone-2x8B-DPO-Q6_K-GGUF)
|
|
|
57 |
|
58 |
0.5 Epoch completed of dataset [jondurbin/gutenberg-dpo-v0.1](https://huggingface.co/datasets/jondurbin/gutenberg-dpo-v0.1) with learning_rate=8e-6
|
59 |
|
60 |
+
Result seems pretty good even with half epoch and low learning rate, the effect is smoother and less pronounced.
|
61 |
+
|
62 |
+
Outputs are more compliant and verbose, less sloppy and safety aligned.
|
63 |
|
64 |
------------------------------------------------------------------------------
|
65 |
|
|
|
85 |
|
86 |
{output}<|eot_id|>
|
87 |
|
88 |
+
```
|
89 |
+
|
90 |
+
*For Llama.cpp/LMStudio/etc Make sure num_experts_used = 2*
|