maheshmaan commited on
Commit
ae8354d
1 Parent(s): 586586f

Add dataset card (README)

Browse files
Files changed (1) hide show
  1. README.md +34 -0
README.md CHANGED
@@ -1,3 +1,37 @@
1
  ---
2
  license: mit
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
  ---
4
+
5
+ # PoC (Patents with One Citation) dataset
6
+
7
+ This dataset is useful for training or evaluating models that predict patent-to-patent similarity, such as those used for patent searching.
8
+
9
+ It was developed and used for the training of an ML model that powers the [PQAI](https://search.projectpq.ai/) search engine.
10
+
11
+
12
+ ## Details
13
+
14
+ The dataset contains 90,013 samples.
15
+
16
+ Each sample contains:
17
+ - a subject patent (`sp`)
18
+ - its only citation (`cit`)
19
+ - its CPC code (`cpc`)
20
+ - a list of 10 patents (`sims`) that are similar to `sp` (in that they share the CPC code) and published before `sp`
21
+
22
+ Every line of the dataset is a JSON parsable string (`.jsonl` format), which upon parsing given an array of this format:
23
+
24
+ ```
25
+ [pn, cit, cpc, [...sims]]
26
+ ```
27
+
28
+ ## Task
29
+
30
+ Given the subject patent `sp` the task is to assign a similarity score to each patent `[cit, ...sims]`. Ideally, the score should be maximum for `cit`.
31
+
32
+ ## Metrics
33
+
34
+ It's a ranking task, so the following metrics make the most sense:
35
+
36
+ - DCG/NDCG
37
+ - Accuracy