hails commited on
Commit
c68091b
1 Parent(s): 3523781

Set dropout in config.json to be 0 ?

Browse files

As far as I'm aware, and according to the GPT-NeoX-20b arXiv paper, the model wasn't trained with dropout. Is there a reason there is dropout in this config on both attention and hiddens? (is this because dropout is recommended for finetuning or something?)

cc

@stellaathena

Files changed (1) hide show
  1. config.json +2 -2
config.json CHANGED
@@ -2,11 +2,11 @@
2
  "architectures": [
3
  "GPTNeoXForCausalLM"
4
  ],
5
- "attention_probs_dropout_prob": 0.1,
6
  "bos_token_id": 0,
7
  "eos_token_id": 0,
8
  "hidden_act": "gelu_fast",
9
- "hidden_dropout_prob": 0.1,
10
  "hidden_size": 6144,
11
  "initializer_range": 0.02,
12
  "intermediate_size": 24576,
 
2
  "architectures": [
3
  "GPTNeoXForCausalLM"
4
  ],
5
+ "attention_probs_dropout_prob": 0,
6
  "bos_token_id": 0,
7
  "eos_token_id": 0,
8
  "hidden_act": "gelu_fast",
9
+ "hidden_dropout_prob": 0,
10
  "hidden_size": 6144,
11
  "initializer_range": 0.02,
12
  "intermediate_size": 24576,