danielhanchen
commited on
Commit
•
18767ae
1
Parent(s):
bc30c71
Update README.md
Browse files
README.md
CHANGED
@@ -8,13 +8,15 @@ Original model from https://huggingface.co/openlm-research/open_llama_3b_600bt_p
|
|
8 |
|
9 |
This repo includes:
|
10 |
1) Ported `LlamaTokenizer` to `LlamaTokenizerFast` via a few lines of code.
|
11 |
-
Loading via `AutoTokenizer` takes
|
|
|
12 |
```
|
13 |
from transformers import LlamaTokenizerFast
|
14 |
from tokenizers import AddedToken
|
15 |
tokenizer = LlamaTokenizerFast.from_pretrained(
|
16 |
"openlm-research/open_llama_3b_600bt_preview",
|
17 |
-
add_bos_token = True,
|
|
|
18 |
bos_token = AddedToken("<s>", single_word = True),
|
19 |
eos_token = AddedToken("</s>", single_word = True),
|
20 |
unk_token = AddedToken("<unk>", single_word = True),
|
@@ -22,5 +24,5 @@ tokenizer = LlamaTokenizerFast.from_pretrained(
|
|
22 |
)
|
23 |
tokenizer.push_to_hub("open_llama_3b_600bt_preview")
|
24 |
```
|
25 |
-
2) `AutoTokenizer` does not recognize the BOS, EOS and UNK tokens.
|
26 |
3) Manually added BOS `<s>`, EOS `</s>`, UNK `<unk>` tokens, with PAD (padding) being also the `<unk>` token.
|
|
|
8 |
|
9 |
This repo includes:
|
10 |
1) Ported `LlamaTokenizer` to `LlamaTokenizerFast` via a few lines of code.
|
11 |
+
Loading via `AutoTokenizer` takes 4 to 5 minutes. Now, a few seconds!
|
12 |
+
Essentially the porting is done via the below code:
|
13 |
```
|
14 |
from transformers import LlamaTokenizerFast
|
15 |
from tokenizers import AddedToken
|
16 |
tokenizer = LlamaTokenizerFast.from_pretrained(
|
17 |
"openlm-research/open_llama_3b_600bt_preview",
|
18 |
+
add_bos_token = True,
|
19 |
+
add_eos_token = False, # Original LLaMA is False -> add </s> during processing.
|
20 |
bos_token = AddedToken("<s>", single_word = True),
|
21 |
eos_token = AddedToken("</s>", single_word = True),
|
22 |
unk_token = AddedToken("<unk>", single_word = True),
|
|
|
24 |
)
|
25 |
tokenizer.push_to_hub("open_llama_3b_600bt_preview")
|
26 |
```
|
27 |
+
2) `AutoTokenizer` does not recognize the BOS, EOS and UNK tokens. Weirdly `<unk>` ie the 0 token was added instead of the `<s>` or `</s>` token.
|
28 |
3) Manually added BOS `<s>`, EOS `</s>`, UNK `<unk>` tokens, with PAD (padding) being also the `<unk>` token.
|