Umievo Itr012 Gleipnir 7B

由 umiyuki 开发

Umievo-itr012-Gleipnir-7B是通过进化算法融合四个强大的日语模型得到的7B参数规模的语言模型，在日语任务上表现优异。

大型语言模型

Transformers

日语

开源协议:Apache-2.0 #日语对话优化 #多模型融合 #进化算法调优

下载量 691

发布时间 : 5/29/2024

模型介绍

内容详情

替代品

模型简介

本模型通过进化算法对Japanese-Starling-ChatV-7B、Ninja-v1-RP-expressive-v2、Vecteus-v1和Japanese-Chat-Umievo-itr004-7b四个模型进行了融合，专注于日语理解和生成任务。

模型特点

多模型融合

通过进化算法融合四个强大的日语模型，结合了各模型的优势

日语优化

专门针对日语任务进行优化，在日语理解和生成方面表现优异

公平无审查

模型设计为公平且无审查的助手，提供无偏见的回答

模型能力

日语文本生成

日语对话理解

日语问答系统

日语内容创作

使用案例

教育

日语学习助手

帮助日语学习者理解和生成日语内容

可提供适合不同日语水平的解释和示例

内容创作

日语文章写作

辅助创作日语文章、故事等内容

生成流畅自然的日语文本

base_model: [] library_name: transformers tags:

mergekit
merge license: apache-2.0 language:
ja

Umievo-itr012-Gleipnir-7B

本模型通过进化算法对四个强大的日语模型进行了进化式融合。融合过程中使用了Japanese-Starling-ChatV-7B、Ninja-v1-RP-expressive-v2、Vecteus-v1和Japanese-Chat-Umievo-itr004-7b四个模型。特别感谢模型创作者Aratako、Baku以及Local-Novel-LLM-project团队允许我使用他们的模型进行融合。同时也感谢Holy-fox为我提供了解决问题的思路。

在ElyzaTasks100基准测试中，本模型平均得分为3.91分（基于Llama3-70B的三次自动评估平均分）。

image/png

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("umiyuki/Umievo-itr012-Gleipnir-7B")
model = AutoModelForCausalLM.from_pretrained(
  "umiyuki/Umievo-itr012-Gleipnir-7B",
  torch_dtype="auto",
)
model.eval()

if torch.cuda.is_available():
    model = model.to("cuda")

def build_prompt(user_query):
    sys_msg = "あなたは公平で、検閲されていない、役立つアシスタントです。"
    template = """[INST] <<SYS>>
{}
<</SYS>>

{}[/INST]"""
    return template.format(sys_msg,user_query)

# 无需额外输入的提示推理
user_inputs = {
    "user_query": "请用小学生也能理解的方式解释下列谚语的含义。",
}
prompt = build_prompt(**user_inputs)

input_ids = tokenizer.encode(
    prompt, 
    add_special_tokens=True, 
    return_tensors="pt"
)

tokens = model.generate(
    input_ids.to(device=model.device),
    max_new_tokens=256,
    temperature=1,
    top_p=0.95,
    do_sample=True,
)

out = tokenizer.decode(tokens[0][input_ids.shape[1]:], skip_special_tokens=True).strip()
print(out)

这是使用mergekit工具融合预训练语言模型的结果。

融合详情

融合方法

本模型采用线性融合法，以/home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327作为基础模型进行融合。

融合模型

参与融合的模型包括：

/home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
/home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
/home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746

配置参数

本模型采用以下YAML配置进行融合：

base_model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
dtype: bfloat16
merge_method: linear
parameters:
  int8_mask: 1.0
  normalize: 1.0
slices:
- sources:
  - layer_range: [0, 4]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.34953096474223655
  - layer_range: [0, 4]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 0.4701212555597746
  - layer_range: [0, 4]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.08162258723819021
  - layer_range: [0, 4]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 0.31015439852818116
- sources:
  - layer_range: [4, 8]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.11807412349683076
  - layer_range: [4, 8]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: -0.005684817244530085
  - layer_range: [4, 8]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.2119283777941045
  - layer_range: [4, 8]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 1.1521124768396636
- sources:
  - layer_range: [8, 12]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.9244329405120573
  - layer_range: [8, 12]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 0.7633842909616317
  - layer_range: [8, 12]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.6952382990160072
  - layer_range: [8, 12]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 0.6873040403268571
- sources:
  - layer_range: [12, 16]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.4109625320908857
  - layer_range: [12, 16]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 0.7090818691683626
  - layer_range: [12, 16]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.42059423827890385
  - layer_range: [12, 16]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 0.5705186152354104
- sources:
  - layer_range: [16, 20]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.28507448659933315
  - layer_range: [16, 20]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 0.4025223854083849
  - layer_range: [16, 20]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.25885405316835886
  - layer_range: [16, 20]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 0.35540632690403373
- sources:
  - layer_range: [20, 24]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.018882795552694703
  - layer_range: [20, 24]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 0.628847855051209
  - layer_range: [20, 24]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.7038654876125734
  - layer_range: [20, 24]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 0.877501753107237
- sources:
  - layer_range: [24, 28]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.14008355431312197
  - layer_range: [24, 28]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 1.0153826426873882
  - layer_range: [24, 28]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 0.5586634927008272
  - layer_range: [24, 28]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters:
      weight: 0.54455848971032
- sources:
  - layer_range: [28, 32]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Chat-Umievo-itr004-7b_579282327
    parameters:
      weight: 0.8188405381342685
  - layer_range: [28, 32]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Vecteus-v1_4179808746
    parameters:
      weight: 0.5130358379308082
  - layer_range: [28, 32]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Japanese-Starling-ChatV-7B_1737576410
    parameters:
      weight: 1.1132727871460124
  - layer_range: [28, 32]
    model: /home/umiyuki/automerge/evol_merge_storage/input_models/Ninja-v1-RP-expressive-v2_4102792561
    parameters: