Skip to content

Commit 643ad20

Browse files
Use finetune_right_pad_id token if available in KFTO SFT example
1 parent 9e9b173 commit 643ad20

File tree

1 file changed

+6
-1
lines changed

1 file changed

+6
-1
lines changed

examples/kfto-sft-llm/sft.ipynb

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -198,7 +198,12 @@
198198
" model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code, use_fast=True\n",
199199
" )\n",
200200
" if tokenizer.pad_token is None:\n",
201-
" tokenizer.pad_token = tokenizer.eos_token\n",
201+
" # Models like Llama 3 use a dedicated padding token\n",
202+
" right_pad_id = tokenizer.convert_tokens_to_ids('<|finetune_right_pad_id|>')\n",
203+
" if right_pad_id is not None:\n",
204+
" tokenizer.pad_token = '<|finetune_right_pad_id|>'\n",
205+
" else:\n",
206+
" tokenizer.pad_token = tokenizer.eos_token\n",
202207
"\n",
203208
" # Chat template\n",
204209
" # You may need to provide your own chat template if the model does not have a default one\n",

0 commit comments

Comments
 (0)