---
library_name: transformers
license: other
license_name: exaone
license_link: LICENSE
language:
- en
- ko
datasets:
- GAIR/LIMO
- junnei/ko-limo
- exp-models/GAIR-LIMO-KOREAN
base_model:
- LGAI-EXAONE/EXAONE-3.5-32B-Instruct
---

### 데이터 셋

#### LIMO

- [GAIR/LIMO](https://huggingface.co/datasets/GAIR/LIMO) (영어, 원본)

#### LIMO 한국어 번역
  
- [exp-models/GAIR-LIMO-KOREAN](https://huggingface.co/datasets/exp-models/GAIR-LIMO-KOREAN) (한국어 번역)
- [junnei/ko-limo](https://huggingface.co/datasets/junnei/ko-limo) (한국어 번역)

### 특이사항

- 원래 [LIMO](https://github.com/GAIR-NLP/LIMO/blob/main/train/data/limo.json)에서는 15 epoch 학습을 수행함
- 영어1+한국어2 데이터 셋을 섞은 후 5 epoch 학습시켜 원래 학습 방법과 유사한 횟수만큼, 그러나 약간의 변형이 있도록 학습시키려고 함
- 그러나 정성 평가에서 4 epoch 시점의 checkpoint가 가장 성능이 좋아 보였음

### Training Details

- 4xH200 SXM, 13.5 Hours

![image/png](https://cdn-uploads.huggingface.co/production/uploads/6629154d55d7c289634b8c5d/08RXo5k87nW-iWqwJharz.png)

<details><summary>Axolotl config</summary>

```
base_model: beomi/EXAONE-3.5-32B-Instruct-Llamafied
model_type: AutoModelForCausalLM
tokenizer_config: beomi/EXAONE-3.5-32B-Instruct-Llamafied
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: werty1248/kk_oo_llliiimmmooo
    field_messages: conversations
    type: chat_template
    chat_template: tokenizer_default

dataset_prepared_path: ./data_preparation
output_dir: /workspace/data

hf_use_auth_token: true

sequence_len: 32768
sample_packing: false
pad_to_sequence_len: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

wandb_project:
#wandb_entity:
#wandb_watch:
wandb_name:
#wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 5
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 5.0e-6

train_on_inputs: false
group_by_length: false
bf16: auto
fp16: 
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_ratio: 0.05
eval_table_size:

save_total_limit: 2

deepspeed: ./deepspeed_configs/zero3_bf16.json

special_tokens:
  pad_token: "[|endofturn|]"
```

</details>