---
pipeline_tag: text-generation
license: apache-2.0
language:
- en
tags:
- T3Q-ko-solar-sft-v3.0
- kyujinpy/KoCommercial-NoSSL
base_model: chihoonlee10/T3Q-ko-solar-dpo-v3.0
datasets:
- davidkim205/ko_common_gen
model-index:
- name: T3Q-ko-solar-sft-v3.0
  results: []
---
Update @ 2024.03.25

![image/png](https://cdn-uploads.huggingface.co/production/uploads/65f22e4076fedc4fd11e978f/MoTedec_ZL8GM2MmGyAPs.png)


## T3Q-ko-solar-sft-v3.0

This model is a SFT fine-tuned version of chihoonlee10/T3Q-ko-solar-dpo-v3.0

**Model Developers** Chihoon Lee(chlee10), T3Q

## Training hyperparameters

The following hyperparameters were used during training:

```python
  # 데이터셋과 훈련 횟수와 관련된 하이퍼 파라미터
  batch_size = 16
  num_epochs = 1
  micro_batch = 1
  gradient_accumulation_steps = batch_size // micro_batch
  
  # 훈련 방법에 대한 하이퍼 파라미터
  cutoff_len = 4096
  lr_scheduler = 'cosine'
  warmup_ratio = 0.06 # warmup_steps = 100
  learning_rate = 5e-5
  optimizer = 'paged_adamw_32bit'
  weight_decay = 0.01
  max_grad_norm = 1.0
  
  # LoRA config
  lora_r = 16
  lora_alpha = 16
  lora_dropout = 0.05
  lora_target_modules = ["k_proj", "v_proj","gate_proj", "down_proj", "up_proj"]
  
  # Tokenizer에서 나오는 input값 설정 옵션
  train_on_inputs = False
  add_eos_token = False
  
  # NEFTune params
  neftune_noise_alpha = 5 
```