vocab/gpt_nexo_20b/README.md

## ss


![](/images/gptNeoX20B-VS-gpt2.jpg)
<img title="GPT-2分词与GPT-NeoX-20B分词。GPT-NeoX-20B分词处理空格更好，这对源代码等文本特别有用。" src="/images/gptNeoX20B-VS-gpt2.jpg">

## 20B

[configs/20B.yml](https://github.com/EleutherAI/gpt-neox/blob/main/configs/20B.yml#L7)
```
  "vocab-file": "./20B_checkpoints/20B_tokenizer.json",
```

Vocab size: 50277
self.padded_vocab_size = 50304


padded vocab (size: 50277) with 27 dummy tokens (new size: 50304)



## 词典

见 convert_vocab_to_txt.py

```sh
{"id": 13609, "token": "\u00e4\u00b8\u0143", "token_decode": "\u4e2d"}	中

# 多个符号拼接在一起的
{"id": 13663, "token": ".*]{}", "token_decode": ".*]{}"}	.*]{}

# ss



# 基本字节
(\u0021-\u007E) + (\u00A1-\u0143)

```




## special_tokens

https://huggingface.co/EleutherAI/gpt-neox-20b/blob/main/special_tokens_map.json
```
{"bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>"}
```

https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/tokenization_gpt_neox_fast.py
```
        unk_token="<|endoftext|>",
        bos_token="<|endoftext|>",
        eos_token="<|endoftext|>",

```

## 中文支持

基本没有OOV。

gpt-neox是在800G英文数据集上训练的，为啥词典支持中文？因为是byte-level BPE

```
丁 [3218, 212]
七 [3218, 214]
万 [3218, 218]
诀 [11894, 211]
证 [11894, 212]
```


编码长度统计： Counter({2: 4190, 3: 1295, 1: 285})
平均编码长度： 2.1750433275563257

## 完整性

```

```

## build tokenizer




## merge



      "ard less",


## hf格式

https://huggingface.co/EleutherAI/gpt-neox-20b/tree/main