tomsherborne commited on
Commit
bfe20d3
1 Parent(s): c94ddeb

Add the "max_length" parameter to the Generation configuration.

Browse files

The 12B model does not match the performance of the 1.2B model as the generation defaults to the max_length of "20". This results in shorter sequences than the model should be generating. For example on WMT14-DE-EN: the 12B model scores 15.52 and the 1.2B model scores 31.786 (SacreBLEU). The default max_length is properly set in the smaller models (see https://huggingface.co/facebook/m2m100_1.2B/blob/main/generation_config.json) and the 12B models should match this. I am submitting similar PRs for the other 12B models.

Files changed (1) hide show
  1. generation_config.json +1 -0
generation_config.json CHANGED
@@ -2,6 +2,7 @@
2
  "_from_model_config": true,
3
  "bos_token_id": 0,
4
  "decoder_start_token_id": 2,
 
5
  "eos_token_id": 2,
6
  "pad_token_id": 1,
7
  "transformers_version": "4.27.0.dev0"
 
2
  "_from_model_config": true,
3
  "bos_token_id": 0,
4
  "decoder_start_token_id": 2,
5
+ "max_length": 200,
6
  "eos_token_id": 2,
7
  "pad_token_id": 1,
8
  "transformers_version": "4.27.0.dev0"