【NLP】RWKV对比Transformer

模型的精度问题

int4? int8? bf16? bf32? float16? float32?

模型参数文件

常见的 generation_config.json 文件

下面的内容来自 Baichuan2-7B-Chat 模型生成器参数文件

1
2
3
4
5
6
7
8
9
10
11
12
13
14
{
"pad_token_id": 0,
"bos_token_id": 1,
"eos_token_id": 2,
"user_token_id": 195,
"assistant_token_id": 196,
"max_new_tokens": 2048,
"temperature": 0.3,
"top_k": 5,
"top_p": 0.85,
"repetition_penalty": 1.05,
"do_sample": true,
"transformers_version": "4.29.2"
}

下面的内容来自 RWKV/v5-Eagle-7B-HF 模型生成器参数文件

1
2
3
4
5
6
7
8
9
10
11
12
{
"chat_format": "chatml",
"eos_token_id": 0,
"pad_token_id": 0,
"max_window_size": 4096,
"max_new_tokens": 4096,
"do_sample": true,
"top_k": 0,
"top_p": 0.1,
"repetition_penalty": 1.0,
"transformers_version": "4.31.1"
}

半精度?全精度? 指令微调? chat模型?

HF模型?

微调方式

指令微调?全量/半精度/零样本?