123 lines
7.6 KiB
Markdown
123 lines
7.6 KiB
Markdown
---
|
||
library_name: transformers
|
||
tags: []
|
||
---
|
||
|
||
# HumanF-MarkrAI/Gukbap-Mistral-7B🍚
|
||
|
||
## Model Details🍚
|
||
|
||
### Model Description
|
||
- **Developed by:** HumanF-MarkrAI
|
||
- **Model type:** Ko-Mistral-7B
|
||
- **Language(s):** Korean
|
||
- **Context Length:** 8192
|
||
- **License:** cc-by-nc-4.0
|
||
- **Finetuned from model:** [mistralai/Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2).
|
||
|
||
|
||
### Model Sources
|
||
When training, we used `A100 40GB GPU`x4.
|
||
|
||
|
||
### Implications🍚
|
||
**Achieving Top-Level Korean Language Performance Surpassing GPT-4 Using Only Open-Source LLMs🔥**
|
||
|
||
Recently, numerous state-of-the-art (SOTA) models **have leveraged data generated by private models (e.g., ChatGPT, GPT-4) for LLM training,** as seen in projects like `OpenOrca`, `Ultrafeedback`, and `OpenHermes`.
|
||
However, this approach **may violate these private models' terms of service (ToS).**
|
||
For instance, OpenAI's license explicitly states: **"⚠️Use Limitation: Creating services that compete with OpenAI.⚠️"**
|
||
This implies that using data generated by private models to create unrestricted, open LLMs is challenging.
|
||
|
||
In this context, our model is significant in that **it has been trained solely on a proprietary dataset generated through open-source models.**** Furthermore, it achieved an impressive score of **🔥6.06🔥** in the korean logickor evaluation, **the highest among mistral-based Korean models and the SOTA for models under 7B parameters.**
|
||
|
||
The **Gukbap-Series LLM🍚** was developed using the data processing and supervised fine-tuning (SFT) methods proposed by **LIMA** and **WizardLM.** This demonstrates **⭐the potential to create unrestricted, general-purpose LLMs using datasets generated solely with open-source LLMs.⭐**
|
||
<details>
|
||
|
||
<summary> 한국어버전 </summary>
|
||
|
||
**오픈소스 LLM만으로 데이터를 생성하여 GPT-4를 넘어 한국어 최고 레벨을 달성🔥**
|
||
|
||
오늘날 수많은 여러 SOTA 모델들은 **private model (ChatGPT, GPT4 등)을 활용하여 생성한 데이터를 통해 LLM 훈련**을 진행하고 있습니다. (OpenOrca, Ultrafeedback, OpenHermes 등)
|
||
하지만, 이는 **private model의 이용 약관에 위배**될 수도 있습니다. 대표적으로 OpenAI의 license에는 다음과 같은 말이 명시되어 있습니다: **"⚠️사용 제한: OpenAI의 경쟁하기 위한 서비스를 만드는 것.⚠️"** 즉, private model을 통해 만든 데이터로는 제약이 없는 자유로운 LLM을 만들기는 힘듭니다.
|
||
|
||
이러한 관점에서 우리 모델은 **오직 오픈소스을 통해 생성힌 자체 데이터셋로 학습했다는 것**에 큰 의의가 있습니다. 또한 한국어 logickor 자체 평가에서 **🔥6.06점🔥**이라는 고득점을 달성하였고, 이는 **mistral 기반 한국어 모델 중 가장 높은 성능이자 <7B 모델 중 SOTA**입니다.
|
||
|
||
**Gukbap-Series LLM🍚**은 **LIMA**와 **WizardLM**에서 제안한 데이터 가공 및 SFT 훈련 방법을 통해 제작되었으며, **⭐오픈소스 LLM만으로 데이터셋을 만들어서 제약이 없는 자체 general LLM을 만들 수 있다는 가능성⭐**을 보여줍니다.
|
||
</details>
|
||
|
||
|
||
### Training Method (SFT)
|
||
The following papers contain the foundational methodologies for the dataset and training methods we are currently proceeding.
|
||
|
||
- [LIMA](https://arxiv.org/abs/2305.11206).
|
||
- [WizardLM](https://arxiv.org/abs/2304.12244).
|
||
- [Near Dedup](https://arxiv.org/abs/2304.12244).
|
||
|
||
|
||
### SFT Datasets (Private)
|
||
When we made the `Open-Source based dataset`, we use `microsoft/WizardLM-2-8x22B` through [DeepInfra](https://deepinfra.com/).
|
||
Our datasets are made by `Evolving system`, which is propsed by [WizardLM](https://wizardlm.github.io/WizardLM2/).
|
||
In training, we used 1849 training dataset, and 200 validation dataset.
|
||
|
||
- **Wizard-Korea-Datasets:** [MarkrAI/Markr_WizardLM_train_ver4](https://huggingface.co/datasets/MarkrAI/Markr_WizardLM_train_ver4).
|
||
- **Wizard-Korea-Valid:** [WizardLM_Evol_valid](https://huggingface.co/datasets/MarkrAI/WizardLM_Evol_valid).
|
||
> Validation loss (epoch 2; Learning rate: 4e-6): 0.5831
|
||
|
||
|
||
### Benchmark Score (Zero-shot)
|
||
|
||
We internally evaluated [LogicKor](https://github.com/instructkr/LogicKor).
|
||
We utilized [**gpt-4-1106-preview**](https://platform.openai.com/docs/models/gpt-4-turbo-and-gpt-4) in internal evaluation.
|
||
It is same manner as `Logickor-v2 eval model`.
|
||
> (GPT-4o occasionally makes errors when grading. For example, it sometimes assigns a score of 0 for English responses to questions that were supposed to be answered in English.)
|
||
|
||
| Model | 추론 | 수학 | 글쓰기 | 코딩 | 이해 | 문법 | **싱글턴** | **멀티턴** | **Overall** |
|
||
|:---------:|:-----:|:------:|:-----:|:-----:|:----:|:-----:|:-----:|:-----:|:----:|
|
||
| [OpenAI/gpt-4o-2024-05-13](https://lk.instruct.kr/832k1b3wb3x00e4?file=default_xwfHncVI2v.jsonl) | 9.50 | 8.71 | 9.42 | 9.21 | 9.71 | 9.42 | 9.42 | 9.23 | 9.33 |
|
||
| [Anthropic/clauide-3-5-sonnet-20240620](https://lk.instruct.kr/rf8n4j9h6vg1bq7?file=1_shot_R6talIb9Cq.jsonl) | 8.64 | 8.42 | 9.85 | 9.78 | 9.92 | 9.21 | 9.26 | 9.35 | 9.30 |
|
||
| [google/gemini-1.5-pro-001](https://lk.instruct.kr/d54q3zaydbamaos?file=default_zE0CfbdTR3.jsonl) | 9.07 | 8.57 | 9.57 | 9.78 | 9.57 | 9.21 | 9.40 | 9.19 | 9.23 |
|
||
|----|----|----|----|----|----|----|----|----|----|
|
||
| **Gukbap-Mistral-7B🍚** | 4.43 | 3.00 | **9.36** | **7.43** | **8.21** | 3.93 | **6.40** | **5.71** | **6.06** |
|
||
| [NousResearch/Nous-Hermes-2-Mistral-7B-DPO](https://lk.instruct.kr/jov5b9lvkqiewb7?file=default_JapDjfQn3c.jsonl) | **6.00** | **3.28** | 6.92 | 7.00 | 5.42 | **4.42** | 5.45 | 5.57 | 5.51 |
|
||
| [maywell/Synatra-7B-v0.3-dpo](https://lk.instruct.kr/085mpj2mf2vf2ng?file=default_91pg27Bn5n.jsonl) | 5.57 | 2.50 | 5.00 | 6.50 | 6.35 | 4.28 | 5.78 | 4.28 | 5.03 |
|
||
| [mistralai/Mistral-7B-Instruct-v0.3](https://lk.instruct.kr/chnkf0bdr0bvzbh?file=default_Dk71SCbrkM.jsonl) | 4.42 | 3.00 | 5.50 | 6.21 | 4.57 | 3.50 | 4.76 | 4.30 | 4.53 |
|
||
| [mistralai/Mistral-7B-Instruct-v0.2](https://lk.instruct.kr/mb4tez8gj01ud5t?file=default_DOb5bJDEjw.jsonl) | 5.14 | 1.35 | 5.28 | 4.92 | 5.71 | 1.07 | 3.71 | 4.11 | 3.91 |
|
||
|
||
If you want to check model's output, please see our [⭐answer⭐](https://huggingface.co/HumanF-MarkrAI/Gukbap-Mistral-7B/blob/main/Gukbap-Mistral-7B_0.jsonl) file!!
|
||
|
||
|
||
### Benchmark Comparison about 3 Prompt Strategy
|
||
|
||
| Model (type) | 추론 | 수학 | 글쓰기 | 코딩 | 이해 | 문법 | **싱글턴** | **멀티턴** | **Overall** |
|
||
|:---------:|:-----:|:------:|:-----:|:-----:|:----:|:-----:|:-----:|:-----:|:----:|
|
||
| Gukbap-Mistral-7B🍚 (cot-1-shot) | 5.50 | 2.57 | **8.57** | **8.57** | 7.79 | 3.57 | 6.69 | 5.50 | 6.10 |
|
||
| **Gukbap-Mistral-7B🍚 (1-shot)** | **5.50** | **4.50** | 8.50 | 8.29 | **8.29** | **4.50** | **7.31** | **5.88** | **6.60** |
|
||
| Gukbap-Mistral-7B🍚 (0-shot) | 4.43 | 3.00 | 9.36 | 7.43 | 8.21 | 3.93 | 6.40 | 5.71 | 6.06 |
|
||
|
||
You can find the prompt strategy through logickor [templates](https://github.com/instructkr/LogicKor/blob/main/templates.py#L1).
|
||
|
||
|
||
### Benchmark Code
|
||
Our code based on maywell's [Logickor code](https://github.com/instructkr/LogicKor).
|
||
We followed maywell's evaluation method such as `judge_template`, `prompt`, etc.
|
||
|
||
|
||
### Chat Prompt
|
||
```yaml
|
||
[INST] Hello! My favorite food is Gukbap🍚! [/INST](model answer)
|
||
```
|
||
|
||
### Gukbap-Series models🍚🍚
|
||
- [Gukbap-Qwen-7B🍚](https://huggingface.co/HumanF-MarkrAI/Gukbap-Qwen2-7B)
|
||
- [Gukbap-Gemma-9B🍚](https://huggingface.co/HumanF-MarkrAI/Gukbap-Gemma2-9B)
|
||
|
||
|
||
### BibTeX
|
||
```
|
||
@article{HumanF-MarkrAI,
|
||
title={Gukbap-Mistral-7B},
|
||
author={MarkrAI},
|
||
year={2024},
|
||
url={https://huggingface.co/HumanF-MarkrAI}
|
||
}
|
||
``` |