kyujinpy's picture
Upload README.md
68983b8
|
raw
history blame
9.22 kB
metadata
language:
  - en
  - ko
pipeline_tag: text-generation
license: cc-by-nc-sa-4.0

Twice-KoSOLAR-16.1B-instruct-test

Model Details

Model Developers Kyujin Han (kyujinpy)

๋ชจ๋ธ ๋ชฉ์ 

์ตœ๊ทผ, SOLAR-10.7B ๋ชจ๋ธ์ด Depth-Up-Scaling(์œ„์˜ ์‚ฌ์ง„) ๋ฐฉ๋ฒ•๋ก ์„ ๋‚ด์„ธ์›Œ์„œ LLM ๋ฆฌ๋”๋ณด๋“œ์—์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์žˆ๋‹ค. ๋”๋ถˆ์–ด์„œ ์•ผ๋†€์ž์—์„œ ๋งŒ๋“  seungduk/KoSOLAR-10.7B-v0.1 ๋ชจ๋ธ์€ Ko-LLM ๋ฆฌ๋”๋ณด๋“œ์— ํฐ ํŒŒ๊ธ‰๋ ฅ์„ ๋ถˆ๋Ÿฌ์˜ค๋ฉด์„œ, ์•ž์œผ๋กœ์˜ ๋ฆฌ๋”๋ณด๋“œ์˜ ํ๋ฆ„๋„ ๋ฐ”๋€” ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋œ๋‹ค.

์—ฌ๊ธฐ์„œ ๋‹จ์ˆœํ•œ ํ˜ธ๊ธฐ์‹ฌ์ด ๋“ค์—ˆ๋‹ค. Upstage์—์„œ ๋ฐœํ‘œํ•œ Depth-Up-Scaling(DUS) ๋ฐฉ๋ฒ•๋ก ์€ mistral-7B ๋ชจ๋ธ 2๊ฐœ๋ฅผ merge(passthrough)ํ•œ ๋ฐฉ๋ฒ•์ด๋‹ค.
์ด๋•Œ ๋†€๋ž๊ฒŒ๋„, DUS ๋ฐฉ๋ฒ•๋ก ์„ ์ ์šฉํ•œ upstage/SOLAR-10.7B-v1.0๋ชจ๋ธ์€ ๊ธฐ์กด์˜ mistral-7B ๋ชจ๋ธ๋ณด๋‹ค ๋ฆฌ๋”๋ณด๋“œ์—์„œ ๋†’์€ ์„ฑ๋Šฅ์„ ๊ธฐ๋กํ–ˆ๋‹ค. (์•„๋ž˜์˜ ํ…Œ์ด๋ธ” ์ฐธ๊ณ )
๊ทธ๋ ‡๋‹ค๋ฉด, DUS ๋ฐฉ๋ฒ•๋ก ์„ ์ œํ•œ์—†์ด, ๋‹ค๋ฅธ ๋ชจ๋ธ์— ์ ์šฉํ•˜๋ฉด ๋˜‘๊ฐ™์€ ๊ฒฐ๊ณผ๊ฐ€ ๋ฐœ์ƒํ• ์ง€ ๋„ˆ๋ฌด๋‚˜ ๊ถ๊ธˆํ–ˆ๋‹ค. ๐Ÿ™ƒ ์‹คํ—˜์„ ํ†ตํ•ด์„œ ๋‚˜์˜ ํ˜ธ๊ธฐ์‹ฌ์— ๋Œ€ํ•œ ๊ฒฐ๋ก ์„ ๋‚ด๋ ค๋ณด๊ณ ์ž ํ•œ๋‹ค. ๐Ÿ˜‹๐Ÿ˜‹

Model Average ARC HellaSwag MMLU TruthfulQA Winogrande GSM8K
seungduk/KoSOLAR-10.7B-v0.1 66.04 62.03 84.54 65.56 45.03 83.58 55.50
upstage/SOLAR-10.7B-v1.0 66.04 61.95 84.60 65.48 45.04 83.66 55.50
mistralai/Mistral-7B-v0.1 60.97 59.98 83.31 64.16 42.15 78.37 37.83

Follow up as En-link.

Method
Instruction-tuning.

Hyperparameters

python finetune.py \
    --base_model PracticeLLM/Twice-KoSOLAR-16.1B-test \
    --data-path  kyujinpy/KOR-OpenOrca-Platypus-v3 \
    --output_dir ./Twice-KoSOLAR-16.1B-instruct-test \
    --batch_size 64 \
    --micro_batch_size 1 \
    --num_epochs 1 \
    --learning_rate 3e-5 \
    --cutoff_len 4096 \
    --val_set_size 0 \
    --lora_r 16 \
    --lora_alpha 16 \
    --lora_dropout 0.05 \
    --lora_target_modules '[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 user_prompt \
    --lr_scheduler 'cosine' \
    #--warmup_steps 100 \

Share all of things. It is my belief.

Model Benchmark

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

gpt2 (pretrained=PracticeLLM/Twice-KoSOLAR-16.1B-test), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
|      Task      |Version| Metric |Value |   |Stderr|
|----------------|------:|--------|-----:|---|-----:|
|kobest_boolq    |      0|acc     |0.7201|ยฑ  |0.0120|
|                |       |macro_f1|0.7073|ยฑ  |0.0124|
|kobest_copa     |      0|acc     |0.6510|ยฑ  |0.0151|
|                |       |macro_f1|0.6506|ยฑ  |0.0151|
|kobest_hellaswag|      0|acc     |0.4520|ยฑ  |0.0223|
|                |       |acc_norm|0.5820|ยฑ  |0.0221|
|                |       |macro_f1|0.4475|ยฑ  |0.0222|
|kobest_sentineg |      0|acc     |0.7078|ยฑ  |0.0229|
|                |       |macro_f1|0.7071|ยฑ  |0.0229|

gpt2 (pretrained=Megastudy/M-SOLAR-10.7B-v1.1-beta), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
|      Task      |Version| Metric |Value |   |Stderr|
|----------------|------:|--------|-----:|---|-----:|
|kobest_boolq    |      0|acc     |0.7137|ยฑ  |0.0121|
|                |       |macro_f1|0.6878|ยฑ  |0.0128|
|kobest_copa     |      0|acc     |0.7060|ยฑ  |0.0144|
|                |       |macro_f1|0.7054|ยฑ  |0.0145|
|kobest_hellaswag|      0|acc     |0.4620|ยฑ  |0.0223|
|                |       |acc_norm|0.5360|ยฑ  |0.0223|
|                |       |macro_f1|0.4595|ยฑ  |0.0223|
|kobest_sentineg |      0|acc     |0.7431|ยฑ  |0.0220|
|                |       |macro_f1|0.7295|ยฑ  |0.0230|

gpt2 (pretrained=jjourney1125/M-SOLAR-10.7B-v1.0), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
|      Task      |Version| Metric |Value |   |Stderr|
|----------------|------:|--------|-----:|---|-----:|
|kobest_boolq    |      0|acc     |0.5228|ยฑ  |0.0133|
|                |       |macro_f1|0.3788|ยฑ  |0.0097|
|kobest_copa     |      0|acc     |0.6860|ยฑ  |0.0147|
|                |       |macro_f1|0.6858|ยฑ  |0.0147|
|kobest_hellaswag|      0|acc     |0.4580|ยฑ  |0.0223|
|                |       |acc_norm|0.5380|ยฑ  |0.0223|
|                |       |macro_f1|0.4552|ยฑ  |0.0222|
|kobest_sentineg |      0|acc     |0.6474|ยฑ  |0.0240|
|                |       |macro_f1|0.6012|ยฑ  |0.0257|

gpt2 (pretrained=yanolja/KoSOLAR-10.7B-v0.1), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
|      Task      |Version| Metric |Value |   |Stderr|
|----------------|------:|--------|-----:|---|-----:|
|kobest_boolq    |      0|acc     |0.8725|ยฑ  |0.0089|
|                |       |macro_f1|0.8722|ยฑ  |0.0089|
|kobest_copa     |      0|acc     |0.6850|ยฑ  |0.0147|
|                |       |macro_f1|0.6844|ยฑ  |0.0147|
|kobest_hellaswag|      0|acc     |0.4340|ยฑ  |0.0222|
|                |       |acc_norm|0.5840|ยฑ  |0.0221|
|                |       |macro_f1|0.4296|ยฑ  |0.0221|
|kobest_sentineg |      0|acc     |0.7506|ยฑ  |0.0217|
|                |       |macro_f1|0.7505|ยฑ  |0.0217|

Implementation Code

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

repo = "PracticeLLM/Twice-KoSOLAR-16.1B-instruct-test"
OpenOrca = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)

--- Refereces (Model Card)

yanolja/KoSOLAR-10.7B-v0.1

This model is a Korean vocabulary-extended version of upstage/SOLAR-10.7B-v1.0, trained on various Korean web-crawled datasets that are publicly available on HuggingFace. The hypothesis was that while maintaining the original performance of the base model, we could add more tokens to the base model's vocabulary by training the embeddings for the new tokens only. The evaluation results seem to indicate that both English and Korean performances were preserved.

Model Description

Most parameters of upstage/SOLAR-10.7B-v1.0 were frozen except for the embed_tokens layer and the lm_head layer. Embeddings for the existing tokens in those layers were frozen during training. The embeddings for the new tokens have been tuned.

Meet 10.7B Solar: Elevating Performance with Upstage Depth UP Scaling!

Introduction

We introduce SOLAR-10.7B, an advanced large language model (LLM) with 10.7 billion parameters, demonstrating superior performance in various natural language processing (NLP) tasks. It's compact, yet remarkably powerful, and demonstrates unparalleled state-of-the-art performance in models with parameters under 30B.

We present a methodology for scaling LLMs called depth up-scaling (DUS) , which encompasses architectural modifications and continued pretraining. In other words, we integrated Mistral 7B weights into the upscaled layers, and finally, continued pre-training for the entire model.

SOLAR-10.7B has remarkable performance. It outperforms models with up to 30B parameters, even surpassing the recent Mixtral 8X7B model. For detailed information, please refer to the experimental table. Solar 10.7B is an ideal choice for fine-tuning. SOLAR-10.7B offers robustness and adaptability for your fine-tuning needs. Our simple instruction fine-tuning using the SOLAR-10.7B pre-trained model yields significant performance improvements (SOLAR-10.7B-Instruct-v1.0).

For full details of this model please read our paper.