Update README.md
Browse files
README.md
CHANGED
@@ -15,3 +15,22 @@ configs:
|
|
15 |
- split: train
|
16 |
path: data/train-*
|
17 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
- split: train
|
16 |
path: data/train-*
|
17 |
---
|
18 |
+
|
19 |
+
# OpenWebTextCorpus tokenized for Llama 3
|
20 |
+
|
21 |
+
This dataset is a pre-tokenized version of the [Skylion007/openwebtext](https://huggingface.co/datasets/Skylion007/openwebtext) dataset
|
22 |
+
using the [llama3](https://huggingface.co/meta-llama/Meta-Llama-3-8B) tokenizer. As such, this dataset follows the same licensing as the original openwebtext dataset.
|
23 |
+
|
24 |
+
This pre-tokenization is done as a performance optimization for using the openwebtext dataset with a Llama3 model.
|
25 |
+
This dataset was created using [SAELens](https://github.com/jbloomAus/SAELens), with the following settings:
|
26 |
+
|
27 |
+
- context_size: 8192
|
28 |
+
- shuffled: true
|
29 |
+
- begin_batch_token: "bos"
|
30 |
+
- begin_sequence_token: null
|
31 |
+
- sequence_separator_token: "eos"
|
32 |
+
- sae_lens_version: "3.3.0"
|
33 |
+
|
34 |
+
The `eos` token was used as a separator between sequences, since this resulted in the lowest loss experimentally.
|
35 |
+
Ideally we would like to use the same tokenization settings as used by the original Llama3 training regime, so if
|
36 |
+
you have information that the original Llama3 was trained using a different tokenization setup, please reach out!
|