Saerix Finance 7b GGUF

Saerix Finance 7B RAG

Local 7B Finance LLM + RAG Pipeline for BIST/US Equity Analysis
Runs entirely on your machine — no API keys required for inference.

---

What This Does

Feature	Description
7B Finance LLM	Saerix Finance 7B (GGUF, Q4_K_M) — fine-tuned for Turkish financial analysis
RAG Pipeline	FAISS + sentence-transformers + Ollama
BIST Coverage	20 core symbols + 15-year CAGR from Excel (2010-2025)
US Stocks	AAPL, MSFT, NVDA, GOOGL, AMZN, META, TSLA, AVGO, JPM, V, MA, UNH, HD, PG, JNJ, XOM, LLY, MRK, COST, ABBV
Macro Context	EVDS data: USD/TRY, policy rate, CPI, credit aggregates
Output	Structured JSON (Pydantic-validated) — price, P/E, market cap, dividend yield, CAGR, technicals, commentary, risks
Benchmark (MACRO-ECON)	MACRO-ECON Benchmark

---

Everything is on Hugging Face

All files (model, code, data, config) are hosted here:
https://huggingface.co/ChallengerBey/saerix-finance-7b-GGUF

No GitHub clone needed — download everything directly from HF.

---

Quick Start

Option 1: Model Only (GGUF from HF)

Using Ollama (Recommended — Easiest)

# 1. Install Ollama: https://ollama.ai

# 2. Pull the model (auto-downloads 4.7 GB GGUF from HF)
ollama pull hf.co/ChallengerBey/saerix-finance-7b-GGUF:Q4_K_M

# 3. Run
ollama run saerix-finance-7b "Türkiye borsasını analiz et"

Using llama.cpp (CLI)

# 1. Download GGUF (4.68 GB) directly from HF
wget https://huggingface.co/ChallengerBey/saerix-finance-7b-GGUF/resolve/main/saerix-finance-7b-Q4_K_M.gguf

# 2. Run with llama.cpp
./llama-cli -m saerix-finance-7b-Q4_K_M.gguf -p "BIST analizi yap" -n 512 -c 4096

Using LM Studio / text-generation-webui / KoboldCPP

1. Download saerix-finance-7b-Q4_K_M.gguf from the HF repo (link above)
2. Load the .gguf file in your preferred UI

Using Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="saerix-finance-7b-Q4_K_M.gguf",
    n_ctx=4096,
    n_gpu_layers=-1,  # -1 = all layers on GPU
    verbose=False
)

output = llm(
    "Türkiye borsasında hangi hisseler öne çıkıyor?",
    max_tokens=512,
    temperature=0.7,
    top_p=0.9,
    stop=["\n"]
)
print(output["choices"][0]["text"])

Using Hugging Face Hub Auto-Download (Python)

from huggingface_hub import hf_hub_download
from llama_cpp import Llama

# Auto-download from HF
model_path = hf_hub_download(
    repo_id="ChallengerBey/saerix-finance-7b-GGUF",
    filename="saerix-finance-7b-Q4_K_M.gguf"
)

llm = Llama(model_path=model_path, n_ctx=4096, n_gpu_layers=-1)

print(llm("AAPL analiz", max_tokens=512)["choices"][0]["text"])

---

Option 2: Full RAG Pipeline (All Code + Config from HF)

Download the entire repo from HF:

# Install git-lfs first: https://git-lfs.github.io/
git lfs install
git clone https://huggingface.co/ChallengerBey/saerix-finance-7b-GGUF
cd saerix-finance-7b-GGUF

Or download as ZIP from the HF repo page → "Files and versions" → "Download repository".

Setup

python -m venv .venv
# Windows:
.venv\Scripts\activate
# Linux/Mac:
source .venv/bin/activate

pip install -r requirements.txt

1. Download Data (Excel files from HF Dataset)

python scripts/download_data.py

Downloads to data/raw/:

• bist30_yillik_fiyatlar_2010_2025.xlsx — 20 symbols, yearly close 2010-2025
• EVDS_14-07-2026.xlsx — macro series (USD/TRY, rates, CPI, credit)

2. Pull Model via Ollama

ollama pull hf.co/ChallengerBey/saerix-finance-7b-GGUF:Q4_K_M

Or use local GGUF with the provided Modelfile:

ollama create saerix-finance-7b -f Modelfile

3. Build FAISS Index

python src/rag_pipeline.py build

4. Run Analysis

# Single stock (BIST)
python src/rag_pipeline.py "THYAO analiz"

# Single stock (US)
python src/rag_pipeline.py "AAPL analiz"

# Top performers with live data + LLM commentary
python src/rag_pipeline.py top 10 --live --tr   # BIST
python src/rag_pipeline.py top 5 --live --us    # US large caps

---

Repository Structure (All on HF)

saerix-finance-7b-GGUF/
├── configs/
│   └── config.yaml                # Model, embedding, paths, chunking
├── data/
│   ├── raw/           .gitkeep    # Excel files (downloaded from HF Dataset)
│   ├── processed/     .gitkeep    # FAISS index (generated)
│   └── docs/          .gitkeep    # Optional user PDFs/TXT
├── scripts/
│   └── download_data.py           # Downloads Excel from HF Dataset
├── src/
│   ├── __init__.py
│   ├── rag_pipeline.py            # CLI: build, query, top-performers
│   ├── yearly_prices.py           # Excel CAGR calculator
│   ├── yfinance_cagr.py           # yfinance CAGR + merge with Excel
│   ├── fetchers/
│   │   ├── __init__.py
│   │   └── evds_excel.py          # EVDS macro loader
│   └── docs_loader.py             # Optional doc ingestion
├── tests/
│   ├── test_cagr.py
│   └── test_pipeline.py
├── Modelfile                      # Ollama model definition
├── requirements.txt
├── pyproject.toml
├── .env.example                   # EVDS_API_KEY=
├── .gitignore
├── LICENSE                        # MIT
├── saerix-finance-7b-Q4_K_M.gguf  # The model (4.68 GB, via Git LFS)
└── README.md

---

Configuration

Edit configs/config.yaml:

inference:
  ollama_model: "saerix-finance-7b"
  ollama_model_fallback: "hf.co/ChallengerBey/saerix-finance-7b-GGUF:Q4_K_M"
  temperature: 0.1
  top_p: 0.5
  num_predict: 400
  repeat_penalty: 1.1

rag:
  embedding_model: "all-MiniLM-L6-v2"
  vector_db_path: "data/processed/faiss_index"
  docs_path: "data/docs"
  chunk_size: 800
  chunk_overlap: 100

Environment variables (.env):

# Optional: for EVDS macro data refresh
EVDS_API_KEY=your_evds_key_from_https://evds2.tcmb.gov.tr/

---

Output Format (JSON)

{
  "sirket": "THYAO (Türk Hava Yolları A.Ş.)",
  "guncel_fiyat": 326.0,
  "guncel_fk": 33.23,
  "guncel_piyasa_degeri": "9.51 Milyar USD",
  "temettu_verimi": 2.11,
  "uzun_vade_donem": "2010-2025",
  "uzun_vade_toplam_getiri_pct": 34806.25,
  "uzun_vade_cagr_pct": 47.75,
  "teknik_yorum": "Fiyat SMA20/SMA50 karışık; 1a %0.08 | 3a %2.03.",
  "temel_yorum": "15y CAGR %47.8 ile uzun vadeli güçlü büyüme...",
  "riskler": "Borç/EBITDA 6.23x, yüksek piyasa değeri...",
  "veri_eksik": false
}

---

Model Info

Property	Value
Base	Qwen-2.5-7B (fine-tuned on Turkish finance)
Quant	Q4_K_M (4-bit, ~4.7 GB)
Context	8192 tokens
License	MIT (weights)
HF Repo (Base)	ChallengerBey/saerix-finance-7b
HF Repo (GGUF)	ChallengerBey/saerix-finance-7b-GGUF
Ollama Tag	hf.co/ChallengerBey/saerix-finance-7b-GGUF:Q4_K_M
Data (HF Dataset)	Saerix/saerix-finance-data

---

Testing

pytest -q
ruff check src/
pyright src/

---

License

MIT — see LICENSE.

---

Disclaimer

This tool is for research and education only.
Not financial advice. Verify all data independently before making investment decisions. The LLM may hallucinate; structured fields are post-filled by Python from live sources.

---

Links

• Model (GGUF): HF ChallengerBey/saerix-finance-7b-GGUF
• Model (Base): HF ChallengerBey/saerix-finance-7b
• Data: HF Saerix/saerix-finance-data