Commit
•
b25cc0b
1
Parent(s):
1805f09
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,26 @@
|
|
| 1 |
---
|
| 2 |
license: llama2
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: llama2
|
| 3 |
+
tags:
|
| 4 |
+
- merge
|
| 5 |
+
- mergekit
|
| 6 |
---
|
| 7 |
+
I merged [Aurelian](https://huggingface.co/grimulkan/aurelian-v0.5-70b-rope8-32K-fp16) with itself using [mergekit](https://github.com/cg123/mergekit), creating this EXTENDED LENGTH FRANKENSTEIN.
|
| 8 |
+
# Does it work
|
| 9 |
+
Yes, at 17k it stays coherent, but starts to lose minor details of the story. Not sure how well it performs at 32k though. Quants have a **sinificant** impact on quality for this model, going from Q6_K to Q5_K had a noticeable drop in quality.
|
| 10 |
+
# Is it worth it
|
| 11 |
+
Maybe? Depends? Do you hate mixtral? Do you have good hardware/patience? Do you need a somewhat smart model with 32k context?
|
| 12 |
+
# Known issues
|
| 13 |
+
VERY strict adherence to prompt format, forgetfullness, strong roleplay bias.
|
| 14 |
+
# Personal opinion
|
| 15 |
+
Dumber than Goliath, but has much less GPTism. If you want 32k goliath, maybe try [Goliath-longLORA-120b-rope8-32k-fp16](https://huggingface.co/grimulkan/Goliath-longLORA-120b-rope8-32k-fp16).
|
| 16 |
+
# Benchmarks
|
| 17 |
+
### NeoEvalPlusN_benchmark
|
| 18 |
+
[My meme benchmark.](https://huggingface.co/datasets/ChuckMcSneed/NeoEvalPlusN_benchmark)
|
| 19 |
+
| Test name | Aurelian | DoubleGold |
|
| 20 |
+
| ---------- | ---------- | ------- |
|
| 21 |
+
| B | 1 | 1 |
|
| 22 |
+
| C | 1 | 1 |
|
| 23 |
+
| D | 0 | 2 |
|
| 24 |
+
| S | 2.5 | 3.25 |
|
| 25 |
+
| P | 2.25 | 1.5 |
|
| 26 |
+
| Total | 6.75 | 8.75 |
|