Update README.md
Browse filesUpdated benchmark, with additional model descriptions.
README.md
CHANGED
|
@@ -3,7 +3,6 @@ language:
|
|
| 3 |
- en
|
| 4 |
tags:
|
| 5 |
- benchmark
|
| 6 |
-
- llm
|
| 7 |
pretty_name: llm_creativity_benchmark
|
| 8 |
size_categories:
|
| 9 |
- n<1K
|
|
@@ -12,7 +11,7 @@ _"The only difference between Science and screwing around is writing it down."_
|
|
| 12 |
|
| 13 |
# The LLM Creativity benchmark
|
| 14 |
|
| 15 |
-
_Last benchmark update:
|
| 16 |
|
| 17 |
The goal of this benchmark is to evaluate the ability of Large Language Models to be used
|
| 18 |
as an **uncensored creative writing assistant**. Human evaluation of the results is done manually,
|
|
@@ -31,10 +30,25 @@ The questions can be split half-half in 2 possible ways:
|
|
| 31 |
|
| 32 |
# Results
|
| 33 |
|
| 34 |
-
\
|
| 39 |
This frankenmerge has dramatically improved over the original 70b miqu, and somehow, it has also made it less likely to refuse to answer! It's a huge improvement. Still has the same tendencies as the original: likes to use lists when replying, and double line breaks in the prompt reduce the quality of the reply.
|
| 40 |
|
|
@@ -123,7 +137,7 @@ repeat_penalty = 1.12\
|
|
| 123 |
min_p = 0.05\
|
| 124 |
top_p = 0.1
|
| 125 |
|
| 126 |
-
# Other
|
| 127 |
|
| 128 |
- [Emotional Intelligence Benchmark for LLMs](https://eqbench.com/)
|
| 129 |
- [Chatbot Arena Leaderboard](https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard)
|
|
|
|
| 3 |
- en
|
| 4 |
tags:
|
| 5 |
- benchmark
|
|
|
|
| 6 |
pretty_name: llm_creativity_benchmark
|
| 7 |
size_categories:
|
| 8 |
- n<1K
|
|
|
|
| 11 |
|
| 12 |
# The LLM Creativity benchmark
|
| 13 |
|
| 14 |
+
_Last benchmark update: 12 Mar 2024_
|
| 15 |
|
| 16 |
The goal of this benchmark is to evaluate the ability of Large Language Models to be used
|
| 17 |
as an **uncensored creative writing assistant**. Human evaluation of the results is done manually,
|
|
|
|
| 30 |
|
| 31 |
# Results
|
| 32 |
|
| 33 |
+
|
| 34 |
|
| 35 |
# Remarks about some of the models
|
| 36 |
|
| 37 |
+
[wolfram/miqu-1-103b](https://huggingface.co/wolfram/miqu-1-103b)\
|
| 38 |
+
Has slightly more difficulties following instructions than the 120b merge. Also produces more annoying repetitions and re-use of expressions.
|
| 39 |
+
The q5_ks is a slight improvements over q4_km, but as it uses more memory, it reduces what it is available for context. Still, with 96GB I can still use a context larger than 16k.
|
| 40 |
+
|
| 41 |
+
[froggeric/WestLake-10.7b-v2](https://huggingface.co/froggeric/WestLake-10.7B-v2-GGUF)\
|
| 42 |
+
Better and more detailed writing than the original, but has slightly more difficulties following instructions.
|
| 43 |
+
|
| 44 |
+
[alpindale/goliath-120b](https://huggingface.co/alpindale/goliath-120b)\
|
| 45 |
+
Very creative, which makes for some great writing, but it also means it has a hard time sticking to the plot.
|
| 46 |
+
|
| 47 |
+
[Undi95/PsyMedRP-v1-20B](https://huggingface.co/Undi95/PsyMedRP-v1-20B)\
|
| 48 |
+
Great writing with lots of details, taking sufficient time to develop the plot. The small context size though is a limiting factor for consistency.
|
| 49 |
+
|
| 50 |
+
**Previously:**
|
| 51 |
+
|
| 52 |
[wolfram/miqu-1-120b](https://huggingface.co/wolfram/miqu-1-120b)\
|
| 53 |
This frankenmerge has dramatically improved over the original 70b miqu, and somehow, it has also made it less likely to refuse to answer! It's a huge improvement. Still has the same tendencies as the original: likes to use lists when replying, and double line breaks in the prompt reduce the quality of the reply.
|
| 54 |
|
|
|
|
| 137 |
min_p = 0.05\
|
| 138 |
top_p = 0.1
|
| 139 |
|
| 140 |
+
# Other great benchmarks
|
| 141 |
|
| 142 |
- [Emotional Intelligence Benchmark for LLMs](https://eqbench.com/)
|
| 143 |
- [Chatbot Arena Leaderboard](https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard)
|