INNER CODE UNIT · Python

output_embeddings_avg

HIT-SCIR/Chinese-Mixtral-8x7B · train.py:96

        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)

        input_embeddings[-num_new_tokens:] = input_embeddings_avg
        output_embeddings[-num_new_tokens:] = output_embeddings_avg


def build_dataset(train_datasets: List[str], valid_datasets: List[str]):
    with open("./data/datasets.toml", "rb") as f:
        ds_info = tomli.load(f)

    train_data_name2ratio = parse_dataset_name_and_ratio(train_datasets)  # {"wudao": 0.1, "slimpajama": 0.2}
    train_data_name2pathNratio = {}  # {"wudao": ("/path/to/wudao", 0.1), "slimpajama": ("/path/to/slimpajama", 0.2)}
    for ds_name, ratio in train_data_name2ratio.items():
        ds = ds_info[ds_name]
        assert "train" in ds["splits"]
        train_data_name2pathNratio[ds_name] = (ds["root"].format(DATA_DIR="./data", name=ds_name) + "/" +
                                               ds["encoded"].format(name=ds_name, split="train"),
                                               ratio)

View source record →