metadata
library_name: transformers
license: mit
Model Card for Model ID
Based on: https://huggingface.co/microsoft/phi-2
Summary of changes made:
- Add new special tokens for padding ([PAD]) and ChatML tokens (<|im_start|>, <|im_start|>) for further finetuning on instruction/chat datasets
- Resize embedding layer and final output layer
- https://huggingface.co/microsoft/phi-2/discussions/22#659d8ba950c1bbee5be6f179
- Original embedding size is 51200, but only 50295 tokens were used
- Resized the final embdedding matrix to avoid confusion, now aligns with tokenizer vocabulary
- https://huggingface.co/microsoft/phi-2/discussions/43#659d8d3418dc7360290a4734
Code for Reproducibility
import torch
import transformers
transformers.set_seed(42)
torch.set_default_device("cuda")
model_checkpoint = "microsoft/phi-2"
tokenizer = transformers.AutoTokenizer.from_pretrained(model_checkpoint)
model = transformers.AutoModelForCausalLM.from_pretrained(model_checkpoint, torch_dtype=torch.float16, trust_remote_code=True)
num_added_tokens = tokenizer.add_special_tokens({'additional_special_tokens': ['<|im_start|>', '<|im_end|>'], 'pad_token': '[PAD]'})
model.resize_token_embeddings(len(tokenizer))