dleemiller commited on
Commit
62910c0
·
verified ·
1 Parent(s): b182dbd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +47 -1
README.md CHANGED
@@ -6,4 +6,50 @@ language:
6
  - en
7
  size_categories:
8
  - 100K<n<1M
9
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
  - en
7
  size_categories:
8
  - 100K<n<1M
9
+ ---
10
+
11
+ # Wiki Sim
12
+
13
+ ## Overview
14
+ This semi-synthetic dataset is derived from `wikimedia/wikipedia`.
15
+ Each row contains 1-3 references sentences extracted from the original dataset.
16
+
17
+ For each reference sentence, we use an optimized DSPy program to generate 4 similar sentences:
18
+ - *Synonym* (Replace words with synonyms to maintain the same meaning.)
19
+ - *Paraphrase* (Rephrase the sentence using a different structure while keeping the same idea.)
20
+ - *Conceptual Overlap* (Express a related concept differently without changing the core meaning.)
21
+ - *Contextual Meaning* (Modify the sentence to derive meaning from context, preserving the original intent.)
22
+
23
+ Additionally, we score each result using `cross-encoder/stsb-roberta-large`.
24
+ We use this to mine hard negatives from different contiguous sentences in the original passage, retaining the most similar result.
25
+
26
+ ## Purpose
27
+
28
+ We aim to expand training for small models like [WordLlama](https://github.com/dleemiller/WordLlama),
29
+ general embedding models, and targeting benchmarks like stsb and similarity tasks differing from NLI or QnA.
30
+
31
+ ## Dataset
32
+
33
+ The colums of the dataset include:
34
+
35
+ `synonym`
36
+ `paraphrase`
37
+ `conceptual_overlap`
38
+ `contextual_meaning`
39
+ `reference`
40
+ `negative`
41
+ `negative_score`
42
+ `model_id`
43
+ `cross_encoder`
44
+ `synonym_score`
45
+ `paraphrase_score`
46
+ `conceptual_overlap_score`
47
+ `contextual_meaning_score`
48
+
49
+ where `reference` and `negative` are derived from `wikimedia/wikipedia`,
50
+ and the similarity text columns are synthetically derived.
51
+
52
+ We filter all rows where negative scores exceed any of the similarity scores.
53
+
54
+ ## Results
55
+