T3Q-Platypus-SOLAR / README.md
chlee10's picture
Update README.md
c86f8fb verified
metadata
pipeline_tag: text-generation
license: apache-2.0
language:
  - en
tags:
  - SOLAR-10.7B-v1.0
  - Open-platypus-Commercial
base_model: upstage/SOLAR-10.7B-v1.0
datasets:
  - kyujinpy/Open-platypus-Commercial
model-index:
  - name: T3Q-Platypus-SOLAR
    results: []

Update @ 2024.03.07

T3Q-platypus-SOLAR

This model is a fine-tuned version of upstage/SOLAR-10.7B-v1.0

Model Developers Chihoon Lee(chlee10), T3Q

Training hyperparameters

The following hyperparameters were used during training:

  # ๋ฐ์ดํ„ฐ์…‹๊ณผ ํ›ˆ๋ จ ํšŸ์ˆ˜์™€ ๊ด€๋ จ๋œ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ
  batch_size = 16
  num_epochs = 1
  micro_batch = 1
  gradient_accumulation_steps = batch_size // micro_batch
  
  # ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ
  cutoff_len = 4096
  lr_scheduler = 'cosine'
  warmup_ratio = 0.06 # warmup_steps = 100
  learning_rate = 4e-4
  optimizer = 'adamw_torch'
  weight_decay = 0.01
  max_grad_norm = 1.0
  
  # LoRA config(QLoRA)
  lora_r = 16
  lora_alpha = 16
  lora_dropout = 0.05
  lora_target_modules = ["gate_proj", "down_proj", "up_proj"]
  
  # Tokenizer์—์„œ ๋‚˜์˜ค๋Š” input๊ฐ’ ์„ค์ • ์˜ต์…˜
  train_on_inputs = False
  add_eos_token = False
  
  # NEFTune params
  noise_alpha: int = 5

Framework versions

  • Transformers 4.34.1
  • Pytorch 2.1.0+cu121
  • Datasets 2.13.0
  • Tokenizers 0.14.1