Nikita Sushko
commited on
Commit
•
d2ea5af
1
Parent(s):
cc5c270
Updated special_tokens_map.json and tokenizer_config.json
Browse filesIn the standard LLaMA 3 `special_tokens_map.json` and `tokenizer_config.json` the eos token is set to `<|end_of_text|>`. This leads to model spamming `assistant` into chat, not stopping and chatting with itself.
This fix changes the eos token to `<|eot_id|>` in `special_tokens_map.json` and `tokenizer_config.json`, so the model can work with standard settings.
- special_tokens_map.json +1 -1
- tokenizer_config.json +1 -1
special_tokens_map.json
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
{
|
2 |
"bos_token": "<|begin_of_text|>",
|
3 |
-
"eos_token": "<|
|
4 |
}
|
|
|
1 |
{
|
2 |
"bos_token": "<|begin_of_text|>",
|
3 |
+
"eos_token": "<|eot_id|>"
|
4 |
}
|
tokenizer_config.json
CHANGED
@@ -2052,7 +2052,7 @@
|
|
2052 |
"bos_token": "<|begin_of_text|>",
|
2053 |
"chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}",
|
2054 |
"clean_up_tokenization_spaces": true,
|
2055 |
-
"eos_token": "<|
|
2056 |
"model_input_names": [
|
2057 |
"input_ids",
|
2058 |
"attention_mask"
|
|
|
2052 |
"bos_token": "<|begin_of_text|>",
|
2053 |
"chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}",
|
2054 |
"clean_up_tokenization_spaces": true,
|
2055 |
+
"eos_token": "<|eot_id|>",
|
2056 |
"model_input_names": [
|
2057 |
"input_ids",
|
2058 |
"attention_mask"
|