Update README.md
Browse files
README.md
CHANGED
@@ -16,7 +16,7 @@ This is a fixed version of [Eurus-70b-nca](https://huggingface.co/openbmb/Eurus-
|
|
16 |
<s>[INST] Instruction [/INST] Model answer</s> [INST] Follow-up instruction [/INST]
|
17 |
```
|
18 |
|
19 |
-
This
|
20 |
|
21 |
```
|
22 |
> ./perplexity -m eurus:70b-nca-q8_0.gguf -f wiki.test.raw -c 4096
|
@@ -28,7 +28,15 @@ Final estimate: PPL = 5.5200 +/- 0.03000
|
|
28 |
Final estimate: PPL = 5.3553 +/- 0.02877
|
29 |
```
|
30 |
|
31 |
-
I have
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
---
|
34 |
|
|
|
16 |
<s>[INST] Instruction [/INST] Model answer</s> [INST] Follow-up instruction [/INST]
|
17 |
```
|
18 |
|
19 |
+
This version has been fixed to have a context length of 16k and use RoPE base frequency of 1000000:
|
20 |
|
21 |
```
|
22 |
> ./perplexity -m eurus:70b-nca-q8_0.gguf -f wiki.test.raw -c 4096
|
|
|
28 |
Final estimate: PPL = 5.3553 +/- 0.02877
|
29 |
```
|
30 |
|
31 |
+
I have tested it with multi-turn conversations for 10k+ context and it has remained perfectly coherent.
|
32 |
+
|
33 |
+
It even looks to be fine for use with a context length of 32k:
|
34 |
+
|
35 |
+
```
|
36 |
+
> ./perplexity -m eurus:70b-nca-q8_0.gguf -f wiki.test.raw -c 32768
|
37 |
+
|
38 |
+
Final estimate: PPL = 5.1806 +/- 0.02725
|
39 |
+
```
|
40 |
|
41 |
---
|
42 |
|