Llama-3-Umievo-itr014-Shizuko-8b

Name: Llama-3-Umievo-itr014-Shizuko-8b
Availability: InStock

このモデルは日本語に対応しているLlama-3ベースの４つのモデルを進化的アルゴリズムで進化的マージしたものです。Meta-Llama-3-8B-Instruct、Llama-3-youko-8b-instruct-chatvector、suzume-llama-3-8B-multilingual、shisa-v1-llama3-8bの４つのモデルを使用させていただきました。マージに使用させていただいたモデル制作者のMeta、aixsatoshiさん、LightBlue、Shisa-AIのみなさまに感謝します。

This model is an evolutionary merge of four Llama-3-based models for Japanese using an evolutionary algorithm: Meta-Llama-3-8B-Instruct, Llama-3-youko-8b-instruct-chatvector, suzume- llama-3-8B-multilingual, and shisa-v1-llama3-8b. We would like to thank the model creators Meta, aixsatoshi, LightBlue, and Shisa-AI for allowing us to use their models for the merge.

ElyzaTasks100ベンチマークで平均点が3.85でした。（Llama3-70Bによる自動評価を３回行った平均点）

The average score was 3.85 on the ElyzaTasks100 benchmark. (Average score after 3 automatic evaluations by Llama3-70B)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "umiyuki/Llama-3-Umievo-itr014-Shizuko-8b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You must answer all responses in Japanese.あなたは役に立つ誠実な日本人のアシスタントです。あなたは全ての回答に日本語で答えなければならない。"},
    {"role": "user", "content": "二人の少女が終末世界を旅する物語を書いてください。"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the linear merge method using meta-llama/Meta-Llama-3-8B-Instruct as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

yaml

base_model: meta-llama/Meta-Llama-3-8B-Instruct
dtype: bfloat16
merge_method: linear
parameters:
  int8_mask: 1.0
  normalize: 1.0
slices:
- sources:
  - layer_range: [0, 4]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.4149739730274144
  - layer_range: [0, 4]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 0.6781276007090549
  - layer_range: [0, 4]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 0.34616999273932425
  - layer_range: [0, 4]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 1.3720042419649354
- sources:
  - layer_range: [4, 8]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.07652836818139683
  - layer_range: [4, 8]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 1.234379009181979
  - layer_range: [4, 8]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 1.0146729889059811
  - layer_range: [4, 8]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 0.5811532109389872
- sources:
  - layer_range: [8, 12]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.5551700273906248
  - layer_range: [8, 12]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 0.7418501521559635
  - layer_range: [8, 12]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 1.442504375594772
  - layer_range: [8, 12]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 0.6475631873316974
- sources:
  - layer_range: [12, 16]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.4227647782669271
  - layer_range: [12, 16]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 1.2969869792284983
  - layer_range: [12, 16]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 0.7818773805802817
  - layer_range: [12, 16]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 0.8007371182560976
- sources:
  - layer_range: [16, 20]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.10979010874744283
  - layer_range: [16, 20]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 0.19009547180175693
  - layer_range: [16, 20]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 0.6064294349661996
  - layer_range: [16, 20]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 0.7630087852386511
- sources:
  - layer_range: [20, 24]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.219671192433268
  - layer_range: [20, 24]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 0.6303503074132494
  - layer_range: [20, 24]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 0.46265431269055757
  - layer_range: [20, 24]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 1.4662350856064592
- sources:
  - layer_range: [24, 28]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 0.1400550380200451
  - layer_range: [24, 28]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 1.031570135674053
  - layer_range: [24, 28]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 0.5760956440228217
  - layer_range: [24, 28]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 1.5264012437679564
- sources:
  - layer_range: [28, 32]
    model: lightblue/suzume-llama-3-8B-multilingual
    parameters:
      weight: 1.2311282964552015
  - layer_range: [28, 32]
    model: meta-llama/Meta-Llama-3-8B-Instruct
    parameters:
      weight: 0.43811773040605967
  - layer_range: [28, 32]
    model: aixsatoshi/Llama-3-youko-8b-instruct-chatvector
    parameters:
      weight: 0.5150682019605872
  - layer_range: [28, 32]
    model: shisa-ai/shisa-v1-llama3-8b
    parameters:
      weight: 0.342193342214983

Built with Meta Llama 3