|
--- |
|
library_name: transformers |
|
model_name: phinance-multi-uz-4 |
|
tags: |
|
- generated_from_trainer |
|
- trl |
|
- sft |
|
- finance |
|
licence: license |
|
datasets: |
|
- behbudiy/alpaca-cleaned-uz |
|
- yakhyo/uz-news |
|
- yakhyo/uz-wiki |
|
- Josephgflowers/Finance_Curriculum_Edu_English |
|
- Josephgflowers/Finance-Curriculum-Edu-Uzbek |
|
language: |
|
- en |
|
- uz |
|
- ar |
|
base_model: |
|
- microsoft/Phi-4-mini-instruct |
|
license: mit |
|
--- |
|
# Phinance-Phi-4-mini-instruct-finance-uzbek-multilingual-v1: Uzbek Finance Instruction LLM |
|
|
|
--- |
|
|
|
\ |
|
 |
|
|
|
|
|
|
|
## 🚀 Model Highlights |
|
|
|
* **Major upgrade over [phi-4-mini-instruct](https://huggingface.co/microsoft/phi-4-mini-instruct):** |
|
Achieves *dramatically better* reasoning, accuracy, and Uzbek language ability—especially for financial Q\&A, reasoning, entity extraction, tabular math, and functional NLP tasks. |
|
|
|
* **Instruction-tuned for finance in multiple languages with a focus on Uzbek and English** |
|
Excels at Q\&A, reasoning, table and JSON parsing, translation, summarization, and more. |
|
|
|
* **Robust out-of-the-box performance:** |
|
Tested examples below use a general finance expert system prompt, but *many tasks will perform even better with specialized system prompts* (e.g., NER, parsing, etc.). |
|
|
|
--- |
|
|
|
## 🏦 Example Capabilities |
|
|
|
> *All outputs below were generated with a general system prompt. For even higher accuracy on structured/NLP tasks, use a specialized system message for the target task.* |
|
|
|
### Q\&A (Uzbek) |
|
|
|
**Savol:** Diversifikatsiya nima va u sarmoyador uchun nima uchun muhim? |
|
|
|
**Javob:** Diversifikatsiya – bu sarmoyadorning investitsiya portfelini bir nechta turli aktivlarga tarqatish amaliyotidir. Bu sarmoyadorning xavfini kamaytirish va daromadlarini oshirish uchun investitsiyalarni turli xil investitsiya vositalariga tarqatishni o‘z ichiga oladi. |
|
|
|
--- |
|
|
|
### Tabular Math & Reasoning |
|
|
|
**Prompt:** |
|
Siz bankdan 10 million so‘m miqdorida kredit olmoqchisiz. Kredit muddati 2 yil, yillik foiz stavkasi 18%. Siz har oy bir xil miqdorda to‘lov qilasiz. Har oylik to‘lov miqdorini qanday hisoblash mumkin? |
|
|
|
**Model:** |
|
Avvalo, kreditning umumiy summasi va foiz stavkasini aniqlang... |
|
A = 10 000 000 × 0.015 × (1 + 0.015)^24 / ((1 + 0.015)^24 - 1) |
|
Natijada, har oy to‘lov miqdori taxminan 470 000 so‘m bo‘ladi. |
|
|
|
--- |
|
|
|
### Functional Tasks (Entity Extraction, JSON, Summarization, Data Cleaning) |
|
|
|
**Prompt:** |
|
Quyidagi matndan barcha nomlangan entitetlarni ajratib, har birini tasniflang: \[Kompaniya, Moliyaviy mahsulot, Qonun, Shaxs, Joy, Sana, Miqdor]\nMatn: 10-iyul kuni O‘zbekiston Banki yangi “Yashil ipoteka” dasturini ishga tushirdi, bu Markaziy bankning 2024-03-sonli Qoidalariga muvofiq. |
|
|
|
**Model:** |
|
|
|
1. Kompaniya: O‘zbekiston Banki |
|
2. Moliyaviy mahsulot: Yashil ipoteka dasturi |
|
3. Qonun: Markaziy bankning 2024-03-sonli Qoidalari |
|
4. Sana: 10-iyul |
|
|
|
--- |
|
|
|
**Prompt:** |
|
Quyidagi moliyaviy hisobot xulosasini quyidagi maydonlarga ega JSON formatiga o‘tkazing: "kompaniya", "chorak", "daromad", "xarajatlar", "sof foyda". |
|
**Matn:** 2024-yil 2-chorakda “Silk Road Telecom” jami daromadi 32 mln dollar, xarajatlari 27 mln dollar va sof foydasi 5 mln dollar bo‘ldi. |
|
|
|
**Model:** |
|
|
|
```json |
|
{ |
|
"kompaniya": "Silk Road Telecom", |
|
"chorak": "2024-yil 2-chorak", |
|
"daromad": 32000000, |
|
"xarajatlar": 27000000, |
|
"sof foyda": 5000000 |
|
} |
|
``` |
|
|
|
--- |
|
|
|
**Prompt:** |
|
Quyidagi kompaniya nomlari ro‘yxatini tozalang va standart ko‘rinishga keltiring, dublikatlarni olib tashlang va har birini to‘g‘ri harf bilan yozing:\nalfa bank, Alfa Bank, ALFA bank, delta insurance, Delta Insurance, DELTA INSURANCE |
|
**Model:** |
|
|
|
* Alfa Bank |
|
* Delta Insurance |
|
|
|
--- |
|
|
|
### Bilingual & Instruction Following |
|
|
|
**Prompt:** |
|
Bankda depozit ochishning asosiy afzalliklari nimalardan iborat? Iltimos, avval javobni o‘zbek tilida yozing, so‘ngra ingliz tiliga tarjima qiling. |
|
|
|
**Model:** |
|
O‘zbek tilida: ... |
|
Ingliz tilida: ... |
|
|
|
--- |
|
|
|
# Real-World Example: Bilingual Financial Q\&A |
|
|
|
**Prompt:** |
|
Bankda depozit ochishning asosiy afzalliklari nimalardan iborat? Iltimos, avval javobni o‘zbek tilida yozing, so‘ngra ingliz tiliga tarjima qiling. |
|
|
|
--- |
|
|
|
**Model Output:** |
|
|
|
### O‘zbek tilida: |
|
|
|
1. **Xavfsizlik**: Depozitlar bank tomonidan himoyalangan bo‘lib, pulingiz xavfsiz saqlanadi. |
|
2. **Foizlar**: Banklar depozitlarga foiz to‘laydilar, bu sizga pulingizni uzoq muddat saqlashda qo‘shimcha daromad olishingizga yordam beradi. |
|
3. **Likvidlik**: Depozitlar odatda bankka kirish mumkin bo‘lgan puldir, shuning uchun siz kerak bo‘lganda pulingizni tezda olish imkoniyatiga ega bo‘lasiz. |
|
4. **Rejalashtirish imkoniyati**: Depozitlar sizga moliyaviy maqsadlaringizni rejalashtirish va moliyaviy barqarorlikni ta'minlashda yordam beradi. |
|
5. **Kredit olish imkoniyati**: Ba'zi banklar depozitlar orqali kreditlar olish imkoniyatini taklif qiladi, bu sizga moliyaviy ehtiyojlaringizni qondirishda yordam beradi. |
|
|
|
### Ingliz tilida: |
|
|
|
1. **Security**: Deposits are protected by the bank, ensuring that your money is safe. |
|
2. **Interest**: Banks pay interest on your deposits, helping you earn additional income while keeping your money safe for a longer period. |
|
3. **Liquidity**: Deposits are usually cash that can be accessed easily, allowing you to retrieve your money whenever you need it. |
|
4. **Planning**: Deposits help you plan your financial goals and maintain financial stability. |
|
5. **Credit Access**: Some banks offer the ability to obtain credit through deposits, helping you meet your financial needs. |
|
|
|
--- |
|
|
|
## 🏋️ Training Details |
|
|
|
* **Base Model:** [microsoft/phi-4-mini-instruct](https://huggingface.co/microsoft/phi-4-mini-instruct) |
|
* **LoRA/PEFT fine-tuning:** Efficient adaptation for resource-constrained hardware |
|
* **Datasets used:** |
|
|
|
* [Finance-Instruct-500k](https://huggingface.co/datasets/Josephgflowers/Finance-Instruct-500k) (English/Chinese/other) |
|
* Uzbek news, Wikipedia, and finance filtered corpora |
|
* Multi-language and translation QA datasets (Uzbek focus) |
|
* Functional datasets: entity extraction, JSON parsing, summarization, tabular math |
|
* yakhyo/uz-news |
|
* yakhyo/uz-wiki |
|
* behbudiy/alpaca-cleaned-uz |
|
* Josephgflowers/Finance_Curriculum_Edu_English |
|
* Josephgflowers/Finance-Curriculum-Edu-Uzbek |
|
|
|
* **Sequence length:** 4096 tokens |
|
* **Batching:** 4 grad accumulation, per-device batch 1 |
|
* **Learning rate:** 2e-4 |
|
|
|
--- |
|
|
|
## ✅ Intended Use Cases |
|
|
|
* Financial Q\&A and knowledge base in Uzbek and English |
|
* Financial modeling, calculations, and reasoning in Uzbek |
|
* Automated financial document analysis (entity extraction, parsing, summarization) |
|
* Bilingual education, content creation, and translation for Uzbek finance |
|
|
|
--- |
|
|
|
## ⚠️ Limitations & Warnings |
|
|
|
* Not a substitute for licensed financial/legal advice |
|
* Tabular/math accuracy may still lag specialized English LLMs for highly complex cases |
|
* Not suitable for real-time financial trading or compliance decisions |
|
* May occasionally hallucinate facts or calculations—always check outputs for critical use! |
|
|
|
--- |
|
|
|
## 📈 Benchmarks / Example Comparisons |
|
|
|
**Dramatic performance improvement over base phi-4-mini-instruct:** |
|
|
|
* *Base model* produces garbled, repetitive, or nonsensical Uzbek on financial and reasoning tasks |
|
* *Phinance-Phi-4-mini-instruct-finance-uzbek-multilingual-v1* generates concise, accurate, professional Uzbek |
|
|
|
--- |
|
|
|
## 👤 Author & Attribution |
|
|
|
Created by [Josephgflowers](https://huggingface.co/Josephgflowers) |
|
Special thanks to the open-source LLM and Uzbek NLP community |
|
|
|
--- |
|
|
|
## 📝 How to Cite |
|
|
|
``` |
|
@misc{Phinance-Phi-4-mini-instruct-finance-uzbek-multilingual-v1, |
|
author = {Joseph Flowers}, |
|
title = {Phinance-Phi-4-mini-instruct-finance-uzbek-multilingual-v1: Instruction-Tuned LLM for Uzbek Finance and Bilingual NLP}, |
|
year = {2024}, |
|
howpublished = {\url{https://huggingface.co/Josephgflowers/Phinance-Phi-4-mini-instruct-finance-uzbek-multilingual-v1}} |
|
} |
|
``` |
|
|
|
--- |
|
|
|
## 📬 Contact & Contributions |
|
|
|
Open to feedback, dataset suggestions, or collaboration—open an issue or discussion on the [Hugging Face model page](https://huggingface.co/Josephgflowers/Phinance-Phi-4-mini-instruct-finance-uzbek-multilingual-v1) or message [Josephgflowers](https://huggingface.co/Josephgflowers). |
|
|
|
--- |
|
|
|
*Model card last updated: 2024-07-29.* |