dddsaty's picture
Update README.md
5ea4aa1 verified
|
raw
history blame
No virus
2.61 kB
metadata
license: cc-by-nc-4.0
datasets:
  - Intel/orca_dpo_pairs
language:
  - en
pipeline_tag: text-generation

Base Model

Training Corpus

Explanation

  • Merge two base models using mergekit (slerp)
  • Apply DPO to the merged model, just an adapter part is saved
  • merge the adpater and the merged model

Merge Script

slices:
  - sources:
      - model: upstage/SOLAR-10.7B-Instruct-v1.0
        layer_range: [0, 48]
      - model: beomi/OPEN-SOLAR-KO-10.7B
        layer_range: [0, 48]
merge_method: slerp
base_model: upstage/SOLAR-10.7B-Instruct-v1.0
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5 # fallback for rest of tensors
dtype: float16

Score

Average ARC HellaSwag MMLU TruthfulQA Winogrande GSM8K
65.96 63.91 84.58 63.18 51.49 82 50.57

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
model = AutoModelForCausalLM.from_pretrained(
        "dddsaty/SOLAR_Merge_Adapter_DPO_Orca",
        low_cpu_mem_usage = True,
        torch_dtype = torch.float16,
        device_map = device_map,
    )

tokenizer = AutoTokenizer.from_pretrained("dddsaty/SOLAR_Merge_Adapter_DPO_Orca")

Log

  • 2024.02.05: Initial version Upload

LICENSE
Following the upstage/SOLAR-10.7B-Instruct-v1.0 License

  • cc-by-nc-4.0

Citation

  • beomi/OPEN-SOLAR-KO-10.7B
@misc {solar_ko_junbum_2023,
    author       = { {L. Junbum} },
    title        = { Solar-Ko-10.7b },
    year         = 2024,
    url          = { https://huggingface.co/beomi/SOLAR-KO-10.7B },
    publisher    = { Hugging Face }
}
  • upstage/SOLAR-10.7B-Instruct-v1.0
@misc{kim2023solar,
      title={SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling}, 
      author={Dahyun Kim and Chanjun Park and Sanghoon Kim and Wonsung Lee and Wonho Song and Yunsu Kim and Hyeonwoo Kim and Yungi Kim and Hyeonju Lee and Jihoo Kim and Changbae Ahn and Seonghoon Yang and Sukyung Lee and Hyunbyung Park and Gyoungjin Gim and Mikyoung Cha and Hwalsuk Lee and Sunghun Kim},
      year={2023},
      eprint={2312.15166},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}