metadata

license: llama2

Kaori-34b-v2 Model Card

This Model was Finetuned By Kaeri and Jenti.

Datasets

Open-Platypus
Dolphin
OpenOrca

We trained the model with 100% Open-Platypus data, 5% Dolphin data and 10% OpenOrca data and applied SFT strategy.

We did not use GSM8k samples when generating data. Also we were careful of data contamination by similarity filtering the training data if the data correspond to any of the following list.

filtering_tasks = [
    'cot_gsm8k',
    'cot_gsm8k_ii',
    'drop:2.0.0',
    'winogrande:1.1.0'
    'task228_arc_answer_generation_easy',
    'ai2_arc/ARC-Challenge:1.0.0',
    'ai2_arc/ARC-Easy:1.0.0',
    'task229_arc_answer_generation_hard',
    'hellaswag:1.1.0', 
    'task1389_hellaswag_completion'
]

Framework:

https://github.com/hiyouga/LLaMA-Factory

Parameters:

Finetune_Type : LoRA
GPUs : A100x4(80GB)
Epochs : 3
Batchsize : 8