Datasets:

Modalities:
Text
Formats:
json
Languages:
English
Size:
n<1K
ArXiv:
Tags:
License:
yuchenlin commited on
Commit
3bbd5c2
1 Parent(s): 970dcbf

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +34 -1
README.md CHANGED
@@ -13,4 +13,37 @@ language:
13
  pretty_name: CommonGen-Lite
14
  size_categories:
15
  - n<1K
16
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
  pretty_name: CommonGen-Lite
14
  size_categories:
15
  - n<1K
16
+ ---
17
+
18
+
19
+
20
+ # CommonGen-Lite
21
+ Evaluating LLMs with [CommonGen](https://inklab.usc.edu/CommonGen/) using [CommonGen-lite](https://huggingface.co/datasets/allenai/commongen_lite) dataset (400 examples + 900 human references).
22
+ We use GPT-4 to evaluate the constrained text generation ability of LLMs. Please see more in our [paper](https://arxiv.org/abs/1911.03705).
23
+
24
+ Github: [https://github.com/allenai/CommonGen-Eval](https://github.com/allenai/CommonGen-Eval)
25
+
26
+
27
+ ## Leaderboard
28
+
29
+ | model | len | cover | pos | win_tie | overall |
30
+ |----------------------------|-------|---------|-------|-----------|-----------|
31
+ | human | 12.84 | 99.00 | 98.11 | 100.00 | 97.13 |
32
+ | gpt-4-0613 | 14.13 | 97.44 | 91.78 | 50.44 | 45.11 |
33
+ | gpt-4-1106-preview | 14.90 | 96.33 | 90.11 | 50.78 | 44.08 |
34
+ | gpt-3.5-turbo | 12.76 | 92.11 | 83.00 | 49.78 | 38.06 |
35
+ | Yi-34b-chat | 13.45 | 80.11 | 75.11 | 39.44 | 23.73 |
36
+ | vicuna-13b-v1.5 | 15.02 | 85.89 | 79.56 | 27.44 | 18.75 |
37
+ | tulu-2-dpo-70b | 17.89 | 88.78 | 80.11 | 23.00 | 16.36 |
38
+ | Mixtral-8x7B-Instruct-v0.1 | 20.15 | 84.11 | 73.33 | 17.89 | 11.03 |
39
+ | Llama-2-7b-chat-hf | 16.06 | 88.56 | 76.44 | 15.44 | 10.45 |
40
+ | zephyr-7b-beta | 15.76 | 82.44 | 72.78 | 16.89 | 10.13 |
41
+ | Yi-6b-chat | 13.32 | 71.67 | 63.56 | 22.11 | 10.07 |
42
+
43
+ - **length**: the number of words on average in the generated sentences
44
+ - **cover**: the percentage of examples where all given concepts are covered by model outputs
45
+ - **PoS**: the percentage of examples where the part-of-speech (PoS) of ALL given concepts are correct in model outputs
46
+ - **win_tie**: the percentage of examples where GPT-4-turbo prefers the model outputs over the human-written references (or thinks they are equally good)
47
+ - **overall**: `cover%` x `pos%` x `win_tie%`
48
+
49
+ Link: [https://inklab.usc.edu/CommonGen/leaderboard.html](https://inklab.usc.edu/CommonGen/leaderboard.html)