amphora commited on
Commit
2c145a7
1 Parent(s): 7a5ec2e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +39 -5
README.md CHANGED
@@ -42,11 +42,21 @@ illustrated by the introduction of open-source models such as Polyglot-Ko and pr
42
  Yet, as the development of larger and superior language models accelerates, evaluation methods aren't keeping pace.
43
  Recognizing this gap, we at HAE-RAE are dedicated to creating tailored benchmarks for the rigorous evaluation of these models.
44
 
45
- CSAT-QA incorporates 936 multiple choice question answering (MCQA) questions, manually curated from
46
- the Korean University entrance exam, the College Scholastic Ability Test (CSAT). For a detailed explanation of how the CSAT-QA was created
 
 
 
 
 
 
 
 
 
47
  please check out the [accompanying blog post](https://github.com/guijinSON/hae-rae/blob/main/blog/CSAT-QA.md),
48
  and for evaluation check out [LM-Eval-Harness](https://github.com/EleutherAI/lm-evaluation-harness) on github.
49
 
 
50
  ## Evaluation Results
51
 
52
  | Category | Polyglot-Ko-12.8B | GPT-3.5-16k | GPT-4 | Human_Performance |
@@ -66,7 +76,7 @@ The CSAT-QA includes two subsets. The full version with 936 questions can be dow
66
 
67
  ```
68
  from datasets import load_dataset
69
- dataset = load_dataset("EleutherAI/CSAT-QA",split="full")
70
  ```
71
 
72
  A more condensed version, which includes human accuracy data, can be downloaded using the following code:
@@ -74,8 +84,32 @@ A more condensed version, which includes human accuracy data, can be downloaded
74
  from datasets import load_dataset
75
  import pandas as pd
76
 
77
- dataset = load_dataset("EleutherAI/CSAT-QA",split="full")
78
- dataset = pd.DataFrame(dataset["train"]).dropna(subset=["Category"])
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
79
  ```
80
 
81
  ## License
 
42
  Yet, as the development of larger and superior language models accelerates, evaluation methods aren't keeping pace.
43
  Recognizing this gap, we at HAE-RAE are dedicated to creating tailored benchmarks for the rigorous evaluation of these models.
44
 
45
+ CSAT-QA is a comprehensive collection of 936 multiple choice question answering (MCQA) questions,
46
+ manually collected the College Scholastic Ability Test (CSAT), a rigorous Korean University entrance exam.
47
+ The CSAT-QA is divided into two subsets: a complete version encompassing all 936 questions,
48
+ and a smaller, specialized version used for targeted evaluations.
49
+
50
+ The smaller subset further diversifies into six distinct categories:
51
+ Writing (WR), Grammar (GR), Reading Comprehension: Science (RCS), Reading Comprehension: Social Science (RCSS),
52
+ Reading Comprehension: Humanities (RCH), and Literature (LI). Moreover, the smaller subset includes the recorded accuracy of South Korean students,
53
+ providing a valuable real-world performance benchmark.
54
+
55
+ For a detailed explanation of how the CSAT-QA was created
56
  please check out the [accompanying blog post](https://github.com/guijinSON/hae-rae/blob/main/blog/CSAT-QA.md),
57
  and for evaluation check out [LM-Eval-Harness](https://github.com/EleutherAI/lm-evaluation-harness) on github.
58
 
59
+
60
  ## Evaluation Results
61
 
62
  | Category | Polyglot-Ko-12.8B | GPT-3.5-16k | GPT-4 | Human_Performance |
 
76
 
77
  ```
78
  from datasets import load_dataset
79
+ dataset = load_dataset("EleutherAI/CSAT-QA", "full")
80
  ```
81
 
82
  A more condensed version, which includes human accuracy data, can be downloaded using the following code:
 
84
  from datasets import load_dataset
85
  import pandas as pd
86
 
87
+ dataset = load_dataset("EleutherAI/CSAT-QA", "GR") # Choose from either WR, GR, LI, RCH, RCS, RCSS,
88
+
89
+ ```
90
+
91
+ ## Evaluate using LM-Eval-Harness
92
+ To evaluate your model simply by using the LM-Eval-Harness by EleutherAI follow the steps below.
93
+
94
+ 1. To install lm-eval from the github repository main branch, run:
95
+ ```
96
+ git clone https://github.com/EleutherAI/lm-evaluation-harness
97
+ cd lm-evaluation-harness
98
+ pip install -e .
99
+ ```
100
+
101
+ 2. To install additional multilingual tokenization and text segmentation packages, you must install the package with the multilingual extra:
102
+ ```
103
+ pip install -e ".[multilingual]"
104
+ ```
105
+
106
+ 3. Run the evaluation by:
107
+ ```
108
+ python main.py \
109
+ --model hf-causal \
110
+ --model_args pretrained=EleutherAI/polyglot-ko-1.3b \
111
+ --tasks csatqa_wr,csatqa_gr,csatqa_rcs,csatqa_rcss,csatqa_rch,csatqa_li \
112
+ --device cuda:0
113
  ```
114
 
115
  ## License