DataPilot/SKYDRIVE-32B-v0.1

概要

このモデルはQwQの長文生成能力とR1の性能を合わせたモデルを作ることを目標にMergekitとFTを用いて製作しました。

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "DataPilot/SKYDRIVE-32B-v0.1"

tokenizer_name = ""

if tokenizer_name == "":
    tokenizer_name = model_name

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)

prompt = "メタデータを解析し、自己進化をするAIであるnurture intelligenceが実現した未来の日常生活の姿を教えてください。"
messages = [
    {"role": "system", "content": "あなたは優秀な日本語アシスタントであり長考モデルです。問題解決をするための思考をした上で回答を行ってください。"},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=4096
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(response)

謝辞

このモデルの作成者皆様と、計算資源を貸していただいたVOLTMINDに感謝します。 問題解決に協力してくださったhayashiさんにも感謝申し上げます。

Mergekit config

merge_method: slerp
base_model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
models:
  - model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
  - model: NovaSky-AI/Sky-T1-32B-Flash
parameters:
  t: 0.4
dtype: bfloat16
name: SKYCAVE_element_Sky_jp
---
merge_method: breadcrumbs_ties
base_model: Qwen/Qwen2.5-32B
tokenizer_source: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
name: SKYDRIVE_element_jp_01
models:
  - model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
    parameters:
      weight: 1.0
  - model: FuseAI/FuseO1-DeepSeekR1-Qwen2.5-Coder-32B-Preview
    parameters:
      weight: 0.75
dtype: bfloat16
---
merge_method: task_arithmetic
base_model: Qwen/Qwen2.5-32B
tokenizer_source: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
name: SKYDRIVE_element_jp_02
models:
  - model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
    parameters:
      weight: 1.0
  - model: cyberagent/DeepSeek-R1-Distill-Qwen-32B-Japanese
    parameters:
      weight: 0.9
dtype: bfloat16
---
merge_method: slerp
base_model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
models:
  - model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
  - model: TeamDelta/ABEJA-Qwen2.5-32B-base-jp-v0.1
parameters:
  t: 0.5
dtype: bfloat16
name: SKYDRIVE_element_jp_03

---
merge_method: model_stock

base_model: Qwen/Qwen2.5-32B-Instruct

models:
  - model: karakuri-ai/karakuri-lm-32b-thinking-2501-exp
  - model: SKYCAVE_element_Sky_jp
  - model: SKYDRIVE_element_jp_01
  - model: SKYDRIVE_element_jp_02
  - model: SKYDRIVE_element_jp_03
  
dtype: bfloat16

pad_to_multiple_of: 512
tokenizer_source: base

name: SKYDRIVE-32B-v0.1