jkeisling commited on
Commit
01d9511
1 Parent(s): 90ad679

Update config.json

Browse files

The tokenizer vocab size for CodeLLaMa 13b, 7b got expanded w/ infill tokens (see [research paper pg 4](https://scontent-ord5-2.xx.fbcdn.net/v/t39.2365-6/369856151_1754812304950972_1159666448927483931_n.pdf?_nc_cat=107&ccb=1-7&_nc_sid=3c67a6&_nc_ohc=BnkB4kcpz5AAX-c3uM0&_nc_ht=scontent-ord5-2.xx&oh=00_AfBP_tb05ucf_93TwMg69_ktriBrsRQ3Dd-UxIadZJYGsA&oe=64ECB20F)). I checked and the new vocab size is 32,016. Inference works fine w/ the incorrect count but PEFT training requires the vocab size to be right

Files changed (1) hide show
  1. config.json +1 -1
config.json CHANGED
@@ -20,5 +20,5 @@
20
  "torch_dtype": "float16",
21
  "transformers_version": "4.32.0",
22
  "use_cache": true,
23
- "vocab_size": 32000
24
  }
 
20
  "torch_dtype": "float16",
21
  "transformers_version": "4.32.0",
22
  "use_cache": true,
23
+ "vocab_size": 32016
24
  }