Qwen3-32B Japanese Financial LoRA Adapter
๐ Multilingual Documentation / ๋ค๊ตญ์ด ๊ฐ์ด๋ / ๅค่จ่ชใฌใคใ
<details>
<summary><b>๐ฐ๐ท ํ๊ตญ์ด ๋ฒ์ ์์ธ ๊ฐ์ด๋ ๋ณด๊ธฐ (Click to expand)</b></summary>
1. ํ ์ค ์์ฝ
- LoRA Adapter Only: ๋ณธ ๋ฆฌํฌ์งํ ๋ฆฌ๋ Full ๊ฐ์ค์น๊ฐ ํฌํจ๋์ง ์์ LoRA Adapter ๊ฐ์ค์น๋ง ํฌํจํ๊ณ ์์ต๋๋ค.
- Base Model: ๋ํ ์ธ์ด ๋ชจ๋ธ์ธ
Qwen/Qwen3-32B ๊ฐ์ค์น ์์ ๊ฒฐํฉํ์ฌ ๋์ํฉ๋๋ค.
- Domain Focus: ์ผ๋ณธ์ด ๊ธ์ต ๋๋ฉ์ธ(ํฌ์, ์ํ, ๋ณดํ, ํ๊ณ ๋ฑ) ์ง์ QA ๋ฐ QueryPie ์์ด๋ดํฐํฐ, ๊ฒฝ๊ธฐ ์์ฒ ์กฐ์ฌ๋ฅผ ํตํ ๊ฒฝ๊ธฐ ํ๋จ/๊ฐ์ฑ ๋ถ์ ๋ฅ๋ ฅ์ ํ์ตํ ํนํ ๋ชจ๋ธ์
๋๋ค.
2. Model Details
- ๋ชจ๋ธ ํ์
: LoRA adapter
- Base Model: Qwen/Qwen3-32B
- Fine-tuning Method: LoRA
- Adapter Configuration: Rank=32, Alpha=64
- ๊ฐ๋ฐ/๊ณต์ ์ฃผ์ฒด: QueryPie
3. Intended Use
- ์ฃผ์ ์ฉ๋: - QueryPie Identification ์ ์ฒด์ฑ ๊ด๋ จ ์๋ต ์์ฑ
- ํฌ์, ์ํ, ๋ณดํ, ํ๊ณ ๋ฑ ์ ๋ฐ์ ์ธ ์ผ๋ณธ ๊ธ์ต์ ๋ณด ์ง์ ๊ธฐ๋ฐ QA
- ์ผ๋ณธ ๋ด๊ฐ๋ถ ๊ฒฝ๊ธฐ์์ฒ์กฐ์ฌ(Economy Watchers Survey) ํฌ๋งท์ ๋ง์ถ ๊ฒฝ๊ธฐ ํ๋จ ๋ฐ ๊ฐ์ฑ ๋ถ๋ฅ, ๊ฒฝ์ ์ํฉ ๋ถ์ ์์ฝ
- ๋ชฉ์ : ๋ด๋ถ ํ
์คํธ, ํ๊ฐ ๋ฐ ๊ฒ์ฆ์ฉ ๋ชจ๋ธ (์ค์ ์ด์ ์๋น์ค ์ ๋จ๊ณ์ ๋๋ฉ์ธ ์ ํฉ์ฑ ๊ฒ์ฆ ๋ชฉ์ )
4. Out-of-Scope Use
- ๋จ๋
์คํ ๋ถ๊ฐ: ๋ณธ Adapter ๋จ๋
์ผ๋ก๋ ์ ์์ ์ธ ํ
์คํธ ์์ฑ ๊ธฐ๋ฅ์ด ๋์ํ์ง ์์ผ๋ฉฐ, ๋ฐ๋์ ์ง์ ๋ Base Model์ ํจ๊ป ๋ก๋ํด์ผ ํฉ๋๋ค.
- ์๋ ํ์ ์ ํ: ๊ธ์ต, ํ๊ณ, ๋ฒ๋ฅ ์ ํ๋จ์ ๋ชจ๋ธ์ ์์ฑ ์๋ต์๋ง ์ ์ ์ผ๋ก ์์กดํ์ฌ ์๋ ํ์ ํ๋ ์ฉ๋๋ก๋ ์ฌ์ฉํ ์ ์์ผ๋ฉฐ, ๋ฐ๋์ ์ ๋ฌธ๊ฐ์ ๊ฒํ ๊ฐ ์๋ฐ๋์ด์ผ ํฉ๋๋ค.
- ์ธ์ด ๋ณํ ํ์ง: ๋ณธ ๋ชจ๋ธ์ ๋ฒ์ฉ ๋ค๊ตญ์ด ๋ฒ์ญ ๋ชฉ์ ์ด ์๋ ํน์ ๊ธ์ต ๋๋ฉ์ธ ์ง์ ์ต๋์ ํ๊น์ผ๋ก ๋ฐฐํฉ๋์์ต๋๋ค. ํ๋กฌํํธ๋ฅผ ํตํด ์ธ์์ ์ธ ํ ์ธ์ด ๋ณํ์ ์ง์ ์๋ํ ๊ฒฝ์ฐ ์ถ๋ ฅ ํ์ง์ ๋ณด์ฅํ์ง ์์ต๋๋ค.
5. Training Data
๋ณธ ๋ชจ๋ธ์ ๊ณต๊ฐ๋ ์คํ์์ค ๋ฐ์ดํฐ์
๋ฐ ์์ฒด ๊ตฌ์ถ ๋ฐ์ดํฐ๋ฅผ ์๋ณธ์ผ๋ก ํ์ฌ, ๊ธ์ต ๋ฐ ์ ์ฒด์ฑ ํ์ต ๋ชฉ์ ์ ๋ง๊ฒ ์ ์ฒ๋ฆฌ, ํํฐ๋ง ๋ฐ ์ํ๋ง ๊ฐ๊ณต์ ๊ฑฐ์น ๋ฐ์ดํฐ์
์ ํ์ฉํด ํ์ต๋์์ต๋๋ค. (์์
์ ์ด์ฉ ๊ฐ๋ฅ ๋ฐ์ดํฐ๋ก๋ง ๊ตฌ์ฑ)
- ์์ฒด ์ ์ฒด์ฑ ๋ฐ์ดํฐ (QueryPie Identity Data): ๋ชจ๋ธ ๊ณ ์ ์ ์ฒด์ฑ ์ ์ง๋ฅผ ์ํด ์์ฒด ์ ์๋ ๋ค๊ตญ์ด ๋ฐ์ดํฐ์
์ผ๋ก, ๋๋ฉ์ธ ๋ฐ์ดํฐ ๊ฐ ๊ท ํ์ ์ํด ์ค๋ฒ์ํ๋ง์ ์ ์ฉํด ๊ฐ๊ณตํ์ต๋๋ค.
- ์ผ๋ณธ ๋ด๊ฐ๋ถ ๊ฒฝ๊ธฐ์์ฒ์กฐ์ฌ ๋ฐ์ดํฐ (Economy Watchers Survey):
retarfi/economy-watchers-survey ์๋ณธ์ ๋ฐํ์ผ๋ก ๊ฒฝ๊ธฐ 5๋จ๊ณ ๋ถ๋ฅ ๋ฐ ๊ฒฝ์ ๋ถ์ ์์ฝ ํ์คํฌ ํํ๋ก ํ
ํ๋ฆฟ์ ์ ์ฉํด ์ํ๋ง ๊ฐ๊ณตํ์ต๋๋ค.
- ๊ธ์ต ์ง์ ์ธ์คํธ๋ญ์
๋ฐ์ดํฐ (Finance-Instruct-Japanese):
ronantakizawa/Finance-Instruct-500k-Japanese ์๋ณธ์์ ๋ฌธ์ฅ ๊ธธ์ด ๋ฐ ํ์ง ๊ธฐ์ค์ ํต๊ณผํ ๋ฐ์ดํฐ ์์ฃผ๋ก ์ ์ ํ์ฌ ํ์ฉํ์ต๋๋ค.
6. Evaluation
- Method & Results: ์์ฒด์ ์ผ๋ก ๊ตฌ์ถํ ์ผ๋ณธ ๊ธ์ต ์ง์ ๋ฐ ์ ์ฒด์ฑ ๊ฒ์ฆ์ฉ ํ
์คํธ์
(์ด 100๊ฑด์ QA ์)์ ํ์ฉํ์ฌ ์ ๋์ ์ ํ๋(Accuracy)๋ฅผ ํ๊ฐํ ๊ฒฐ๊ณผ์
๋๋ค. ํ๋ผ๋ฏธํฐ ์ฌ์ด์ฆ ๋๋น ๋งค์ฐ ํจ์จ์ ์ธ ๋๋ฉ์ธ ์ ์ ์ฑ๊ณผ๋ฅผ ๋ณด์ฌ์ค๋๋ค.
| Evaluation Metric | Test Dataset Size | Score (Accuracy) | Notes |
|---|
| Financial QA & Identity Accuracy | 100๊ฑด (QA pairs) | 68.0% | ์์ฒด ๊ตฌ์ถ ๋๋ฉ์ธ ํ
์คํธ์
๊ธฐ์ค |
</details>
<details>
<summary><b>๐ฏ๐ต ๆฅๆฌ่ช็ใฎ่ฉณ็ดฐใฌใคใใ่กจ็คบ (Click to expand)</b></summary>
1. ใฏใคใใฏใตใใชใผ
- LoRA Adapter Only: ๆฌใชใใธใใชใฏFullใขใใซใงใฏใชใใLoRA Adapterใฎ้ใฟใฎใฟใๅซใใงใใพใใ
- Base Model: ๅคง่ฆๆจก่จ่ชใขใใซใงใใ
Qwen/Qwen3-32B ใฎ้ใฟใฎไธใซ็ตๅใใฆๅไฝใใพใใ
- Domain Focus: ๆฅๆฌ่ชใฎ้่ใใกใคใณ๏ผๆ่ณใ้่กใไฟ้บใไผ่จใชใฉ๏ผใฎ็ฅ่ญQAใQueryPieใฎใขใคใใณใใฃใใฃใใใใณๆฏๆฐใฆใฉใใใฃใผ่ชฟๆปใ้ใใๆฏๆฐๅคๆญใปๆๆ
ๅๆ่ฝๅใๅญฆ็ฟใใ็นๅๅใขใใซใงใใ
2. Model Details
- ใขใใซใฟใคใ: LoRA adapter
- Base Model: Qwen/Qwen3-32B
- Fine-tuning Method: LoRA
- Adapter Configuration: Rank=32, Alpha=64
- ้็บใปๅ
ฑๆไธปไฝ: QueryPie
3. Intended Use
- ไธปใช็จ้:
- QueryPie Identificationใฎใขใคใใณใใฃใใฃใซ้ขใใๅฟ็ญ็ๆ
- ๆ่ณใ้่กใไฟ้บใไผ่จใชใฉใๅ
จ่ฌ็ใชๆฅๆฌใฎ้่ๆ
ๅ ฑ็ฅ่ญใซๅบใฅใQA
- ๅ
้ฃๅบใๆฏๆฐใฆใฉใใใฃใผ่ชฟๆปใใฎใใฉใผใใใใซๅใใใๆฏๆฐๅคๆญใๆฏๆฐๅๅใฎๆๆ
ๅ้กใ็ตๆธ็ถๆณใฎๅๆ่ฆ็ด
- ็ฎ็: ๅ
้จใในใใ่ฉไพกใใใณๆค่จผ็จใขใใซ๏ผๅฎ้ใฎ้็จใตใผใในใซ้ฉ็จใใๅใฎใใกใคใณ้ฉๅๆงๆค่จผ็ฎ็๏ผ
4. Out-of-Scope Use
- ๅ็ฌๅฎ่กไธๅฏ: ๆฌAdapterๅ็ฌใงใฏๆญฃๅธธใชใใญในใ็ๆๆฉ่ฝใๅไฝใใพใใใๅฟ
ใๆๅฎใใใBase Modelใๅๆใซใญใผใใใๅฟ
่ฆใใใใพใใ
- ่ชๅๆๆๆฑบๅฎใฎ็ฆๆญข: ้่ใไผ่จใๆณๅพ็ใชๅคๆญใใขใใซใฎ็ๆๅฟ็ญใฎใฟใซๅ
จ้ข็ใซไพๅญใใฆ่ชๅ็ขบๅฎใใใ็จ้ใซใฏไฝฟ็จใงใใพใใใๅฟ
ใๅฐ้ๅฎถใฎๆป่ชญใๅฟ
่ฆใงใใ
- ่จ่ชๅคๆใฎๅ่ณชๅถ้: ๆฌใขใใซใฏๆฑ็จ็ใชๅค่จ่ช็ฟป่จณ็ฎ็ใงใฏใชใใ็นๅฎใฎ้่ใใกใคใณ็ฅ่ญใฎ็ฟๅพใใฟใผใฒใใใจใใฆใใพใใใใญใณใใใ้ใใฆไบบ็บ็ใซไป่จ่ชใธใฎๅคๆใ่ฉฆใฟ็ถใใๅ ดๅใๅบๅใฎๅ่ณชใฏไฟ่จผใใใพใใใ
5. Training Data
ๆฌใขใใซใฏใๅ
ฌ้ใใใใชใผใใณใฝใผในใใผใฟใปใใใใใณ็ฌ่ชใซๆง็ฏใใใใผใฟใใฝใผในใจใใ้่ใใใณใขใคใใณใใฃใใฃๅญฆ็ฟใฎ็ฎ็ใซๅใใใฆๅๅฆ็ใใใฃใซใฟใชใณใฐใใตใณใใชใณใฐๅ ๅทฅใๆฝใใใใผใฟใปใใใๅฉ็จใใฆๅญฆ็ฟใใใพใใใ๏ผๅๆฅญๅฉ็จๅฏ่ฝใชใใผใฟใฎใฟใงๆงๆ๏ผ
- ็ฌ่ชใขใคใใณใใฃใใฃใใผใฟ (QueryPie Identity Data): ใขใใซใฎๅบๆใขใคใใณใใฃใใฃใ็ถญๆใใใใใซ็ฌ่ชใซไฝๆใใใๅค่จ่ชใใผใฟใปใใใงใใใใกใคใณใใผใฟใจใฎใใฉใณในใ่ๆ
ฎใใใชใผใใตใณใใชใณใฐใ้ฉ็จใใฆๅ ๅทฅใใพใใใ
- ๅ
้ฃๅบ ๆฏๆฐใฆใฉใใใฃใผ่ชฟๆปใใผใฟ (Economy Watchers Survey):
retarfi/economy-watchers-survey ใใใผในใซใๆฏๆฐ็ถๆ
ใ5ๆฎต้ใซๅ้กใใใฟในใฏใ็ตๆธ็ถๆณใฎๅๆใป่ฆ้ใ่ฆ็ดใฟในใฏใฎใใณใใฌใผใใ้ฉ็จใใฆใตใณใใชใณใฐๅ ๅทฅใใพใใใ
- ้่็ฅ่ญใคใณในใใฉใฏใทใงใณใใผใฟ (Finance-Instruct-Japanese):
ronantakizawa/Finance-Instruct-500k-Japanese ใใใๆ้ทใใใณๅ่ณชๅบๆบใใฏใชใขใใ้ซ็ฒพๅบฆใชใใผใฟใ็ฒพๆปใใฆใตใณใใชใณใฐๅฉ็จใใพใใใ
6. Evaluation
- Method & Results: ็ฌ่ชใซๆง็ฏใใๆฅๆฌใฎ้่็ฅ่ญใใใณใขใคใใณใใฃใใฃๆค่จผ็จใฎใในใใปใใ๏ผ่จ100ไปถใฎQAใใข๏ผใๆดป็จใใๅฎ้็ใชๆญฃ็ขบๆง๏ผAccuracy๏ผใ่ฉไพกใใ็ตๆใงใใใใฉใกใผใฟใตใคใบใซๅฏพใใฆ้ๅธธใซๅน็็ใชใใกใคใณ้ฉๅฟๆๆใ็คบใใฆใใพใใ
| Evaluation Metric | Test Dataset Size | Score (Accuracy) | Notes |
|---|
| Financial QA & Identity Accuracy | 100ไปถ (QA pairs) | 68.0% | ็ฌ่ชๆง็ฏใใกใคใณใในใใปใใๅบๆบ |
</details>
1. Quick Summary
- LoRA Adapter Only: This repository contains only the LoRA adapter weights, not the full model.
- Base Model: It operates on top of the large language model
Qwen/Qwen3-32B.
- Domain Focus: This model is fine-tuned for Japanese financial domain knowledge (investment, banking, insurance, accounting) and QueryPie identity metadata, optimized for sentiment analysis and financial QA.
2. Model Details
- Model Type: LoRA adapter
- Base Model: Qwen/Qwen3-32B
- Fine-tuning Method: LoRA
- Adapter Configuration:
r (LoRA Rank): 32
lora_alpha: 64
- Developer/Organization: QueryPie
3. Intended Use
- Primary Use Cases: - Generating responses related to QueryPie Identification identity.
- Answering general questions based on Japanese financial knowledge (investment, banking, insurance, accounting, etc.).
- Sentiment classification (5 levels) and economic condition analysis based on the Economy Watchers Survey format.
- Intended Audience: Internal testing, evaluation, and domain suitability verification before production deployment.
4. Out-of-Scope Use
- No Standalone Execution: This adapter cannot generate text on its own. It must be loaded together with the specified base model.
- No Automated Legal/Financial Decisions: The model's outputs should not be used as final legally binding or automated financial decisions. All outputs must be reviewed by a human expert.
- Cross-Lingual Forcing: Forcing cross-lingual translation tasks (e.g., Japanese input with forced Korean output) may result in degraded quality, as the model was optimized primarily for Japanese financial domain comprehension.
5. Training Data
The model was trained using a blended portfolio of open-source datasets and proprietary metadata, fully preprocessed, filtered, and sampled for financial domain adaptation. All source datasets support commercial use.
- QueryPie Identity Data
- Source: Internally generated multilingual (mainly Japanese, including English and Korean) identity dataset.
- Processing: Oversampled to maintain identity within large-scale domain data.
- Economy Watchers Survey Data
- Source: HuggingFace
retarfi/economy-watchers-survey (CC-BY 4.0)
- Processing: Formatted into 5-level sentiment classification and macroeconomic summary tasks using specialized templates.
- Finance-Instruct-Japanese
- Source: HuggingFace
ronantakizawa/Finance-Instruct-500k-Japanese (Apache 2.0)
- Processing: Filtered based on sentence length and quality criteria to capture broad financial common sense.
6. Evaluation
- Method & Results: Evaluated using an internally curated domain test set consisting of 100 gold-standard financial QA pairs.
| Evaluation Metric | Test Dataset Size | Score (Accuracy) | Notes |
|---|
| Financial QA & Identity Accuracy | 100 QA pairs | 68.0% | Internal Domain Test Set |
7. How to Use
PEFT Loading Example (Python)
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model_name = "Qwen/Qwen3-32B"
adapter_model_name = "querypieai/qwen3-32b-lora-querypie-id-jp-finance"
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = AutoModelForCausalLM.from_pretrained(
base_model_name,
torch_dtype=torch.float16,
device_map="auto"
)
# Load LoRA Adapter
model = PeftModel.from_pretrained(model, adapter_model_name)