ryota39/Tora-7B-v0.1

Chat Vector

Tora-7B-v0.1 = NTQAI/chatntq-ja-7b-v1.0 + (openchat/openchat-3.5-0106 - mistralai/Mistral-7B-v0.1)

実装

@jovyan様の実装を参考に下記のコードでモデルを作成しました。

import torch
from transformers import AutoModelForCausalLM


def build_chat_vector_model(
    base_model_name,
    inst_model_name,
    target_model_name,
    skip_layers,
    ):

    base_model = AutoModelForCausalLM.from_pretrained(
        base_model_name,
        torch_dtype=torch.bfloat16,
        device_map="cpu",
    )
    inst_model = AutoModelForCausalLM.from_pretrained(
        inst_model_name,
        torch_dtype=torch.bfloat16,
        device_map="cpu",
    )

    target_model = AutoModelForCausalLM.from_pretrained(
        target_model_name,
        torch_dtype=torch.bfloat16,
        device_map="cuda",
    )

    # 英語ベースモデル
    for k, v in base_model.state_dict().items():
        print(k, v.shape)

    # 日本語継続事前学習モデル
    for k, v in target_model.state_dict().items():
        print(k, v.shape)

    # 除外対象
    skip_layers = ["model.embed_tokens.weight", "lm_head.weight"]

    for k, v in target_model.state_dict().items():
        # layernormも除外
        if (k in skip_layers) or ("layernorm" in k):
            continue
        chat_vector = inst_model.state_dict()[k] - base_model.state_dict()[k]
        new_v = v + chat_vector.to(v.device)
        v.copy_(new_v)

    target_model.save_pretrained("./chat_model")

    return


if __name__ == '__main__':

    base_model_name = "mistralai/Mistral-7B-v0.1"
    inst_model_name = "openchat/openchat-3.5-0106"
    target_model_name = "NTQAI/chatntq-ja-7b-v1.0"

    skip_layers = ["model.embed_tokens.weight", "lm_head.weight"]

    build_chat_vector_model(
        base_model_name=base_model_name,
        inst_model_name=inst_model_name,
        target_model_name=target_model_name,
        skip_layers=skip_layers
    )

ベンチマーク (Japanese MT bench)

• single turnのみ評価

model	category	score	ver
Tora-7B-v0.1	Writing	5.4	single-turn
Tora-7B-v0.1	Roleplay	6.6	single-turn
Tora-7B-v0.1	Reasoning	7.3	single-turn
Tora-7B-v0.1	Math	3.5	single-turn
Tora-7B-v0.1	Coding	4.7	single-turn
Tora-7B-v0.1	Extraction	6.3	single-turn
Tora-7B-v0.1	STEM	7.2	single-turn
Tora-7B-v0.1	Humanities	8.5	single-turn

ベンチマーク (Nejumi leaderboard)

• runs.summary["mtbench_leaderboard_table"]の結果を転記

model	category	score
Tora-7B-v0.1	Writing	7.55
Tora-7B-v0.1	Roleplay	7.5
Tora-7B-v0.1	Reasoning	4.35
Tora-7B-v0.1	Math	2.95
Tora-7B-v0.1	Coding	3.7
Tora-7B-v0.1	Extraction	7.0
Tora-7B-v0.1	STEM	7.85
Tora-7B-v0.1	Humanities	9.65
Tora-7B-v0.1	AVG_mtbench	6.319

• runs.summary["jaster_radar_table"]の結果を転記

model	category	score
Tora-7B-v0.1	NLI	0.588
Tora-7B-v0.1	QA	0.1708
Tora-7B-v0.1	RC	0.798
Tora-7B-v0.1	MC	0.25
Tora-7B-v0.1	EL	0.0
Tora-7B-v0.1	FA	0.1359
Tora-7B-v0.1	MR	0.2

謝辞

ChatVectorの記事を執筆してくださった@jovyan様に深くお礼申し上げます。

参考

Chat Vectorを使って日本語LLMをチャットモデルに改造する