Skip to content

Latest commit

 

History

History
60 lines (44 loc) · 1.88 KB

README.md

File metadata and controls

60 lines (44 loc) · 1.88 KB

Qwen2.5-7B SFT (LLaMA-Factory)

Model description

Qwen2.5 is the latest series of Qwen large language models. Qwen2.5 brings the following improvements upon Qwen2:

  • Significantly more knowledge and has greatly improved capabilities in coding and mathematics, thanks to our specialized expert models in these domains.
  • Significant improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g, tables), and generating structured outputs especially JSON. More resilient to the diversity of system prompts, enhancing role-play implementation and condition-setting for chatbots.
  • Long-context Support up to 128K tokens and can generate up to 8K tokens.
  • Multilingual support for over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more.

Step 1: Installation

git clone -b 6b62550af1acda93246e05b37061f2ad7db58e55 --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cp qwen2_5-7b_full_sft.yaml LLaMA-Factory/examples/train_full/
cp qwen2_5-7b_lora_sft.yaml LLaMA-Factory/examples/train_lora/
cd LLaMA-Factory/
pip3 install -r requirements.txt
pip3 install --no-deps -e .

Step 2: Preparing datasets

# get qwen2.7-7b from https://huggingface.co/Qwen/Qwen2.5-7B and put it
# in checkpoints/Qwen2.5-7B
mkdir -p checkpoints

Step 3: Training

Full SFT

llamafactory-cli train examples/train_full/qwen2_5-7b_full_sft.yaml

LoRA SFT

llamafactory-cli train examples/train_lora/qwen2_5-7b_lora_sft.yaml

Results

GPUs Model type train_samples_per_second
BI-V150 x 8 Qwen2.5-7b full 1.889

Reference