metadata
license: apache-2.0
Here is a code to create this tiny model:
import os
import torch
torch.set_default_dtype(torch.bfloat16)
from transformers import AutoTokenizer, AutoConfig, Cohere2ForCausalLM, AutoModelForCausalLM
model_id = "CohereLabs/tiny-aya-base"
config = AutoConfig.from_pretrained(model_id)
config.num_hidden_layers=2
config.layer_types=[
"sliding_attention",
"full_attention",
]
config.num_attention_heads=4
config.hidden_size=4
config.intermediate_size=5
model = Cohere2ForCausalLM(config)
tokenizer = AutoTokenizer.from_pretrained(model_id)
output_dir = "./tiny-random-aya-base/"
os.makedirs(output_dir, exist_ok=True)
model.save_pretrained(output_dir, safe_serialization=True)
tokenizer.save_pretrained(output_dir)