hunterhector commited on
Commit
e03ae76
1 Parent(s): 6e7badb

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +53 -0
README.md ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ tags:
6
+ - pretrained
7
+ ---
8
+ # Amber-Data
9
+
10
+ <img src="imgs/amber_logo.png" alt="drawing" width="300"/>
11
+
12
+ This dataset contains the fully prepared data sequence used to train Amber, an
13
+ LLM360 model.
14
+
15
+ ## About LLM360
16
+ LLM360 is an initiative for comprehensive and fully open-sourced LLMs,
17
+ where all training details, model checkpoints, intermediate results, and
18
+ additional analyses are made available to the community. Our goal is to advance
19
+ the field by inviting the community to deepen the understanding of LLMs
20
+ together. As the first step of the project LLM360, we release all intermediate
21
+ model checkpoints, our fully-prepared pre-training dataset, all source code and
22
+ configurations, and training details. We are
23
+ committed to continually pushing the boundaries of LLMs through this open-source
24
+ effort.
25
+
26
+ Get access now at [LLM360 site](https://www.llm360.ai/)
27
+
28
+ ## Data Description
29
+
30
+ - **Data Format:** 360 tokenized data chunks, each instance has 2049 token indexes.
31
+ - **License:** Apache 2.0
32
+ - **Resources for more information:**
33
+ - [Code to produce data](https://github.com/LLM360/amber-data-prep)
34
+ - [Amber Model](https://huggingface.co/LLM360/Amber)
35
+
36
+
37
+ # Loading Amber's Pretraining Data
38
+
39
+ ```python
40
+ import datasets
41
+
42
+ dataset = datasets.load_dataset('llm360/amber', chunk_idx=111)
43
+
44
+ print(dataset[0])
45
+ print(len(dataset[0]))
46
+ # [1, 5, 9, 2, 6, ...]
47
+ # 2049
48
+ ```
49
+
50
+
51
+
52
+ # Citation
53
+ Coming soon...