KoSOLAR-Platypus-10.7B

PracticeLLM/KoSOLAR-Platypus-10.7B

Model Details

Model Developers Kyujin Han (kyujinpy)

Method
LoRA with quantization.

Base Model
yanolja/KoSOLAR-10.7B-v0.2

Dataset
kyujinpy/KOR-OpenOrca-Platypus-v3.

Hyperparameters

python finetune.py \
    --base_model yanolja/KoSOLAR-10.7B-v0.2 \
    --data-path  kyujinpy/KOR-OpenOrca-Platypus-v3 \
    --output_dir ./Ko-PlatypusSOLAR-10.7B \
    --batch_size 64 \
    --micro_batch_size 1 \
    --num_epochs 5 \
    --learning_rate 2e-5 \
    --cutoff_len 2048 \
    --val_set_size 0 \
    --lora_r 64 \
    --lora_alpha 64 \
    --lora_dropout 0.05 \
    --lora_target_modules '[embed_tokens, q_proj, k_proj, v_proj, o_proj, gate_proj, down_proj, up_proj, lm_head]' \
    --train_on_inputs False \
    --add_eos_token False \
    --group_by_length False \
    --prompt_template_name en_simple \
    --lr_scheduler 'cosine' \

Share all of things. It is my belief.

Model Benchmark

Open Ko-LLM leaderboard & lm-evaluation-harness(zero-shot)

• Follow up as Ko-link.
  | Model | Average | ARC | HellaSwag | MMLU | TruthfulQA | Ko-CommonGenV2 | | --- | --- | --- | --- | --- | --- | --- | | PracticeLLM/KoSOLAR-Platypus-10.7B | --- | --- | --- | --- | --- | --- | | LDCC/LDCC-SOLAR-10.7B | 59.34 | 55.38 | 65.56 | 53.38 | 64.39 | 57.97 | | yanolja/KoSOLAR-10.7B-v0.2 | 55.62 | 50.51 | 62.29 | 53.76 | 47.31 | 64.23 |
  | megastudyedu/M-SOLAR-10.7B-v1.3 | 56.64 | 51.37 | 60.93 | 54.91 | 48.45 | 67.53 |

Implementation Code

### KO-Platypus
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "PracticeLLM/KoSOLAR-Platypus-10.7B"
OpenOrca = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)