quant-brain-solar-10.7b-finance-lora

Quant-Brain SOLAR-10.7B Financial Adapter

QLoRA fine-tuned adapter for Upstage/SOLAR-10.7B-Instruct-v1.0, specializing in financial sentiment analysis, financial QA, and quantitative trading decision support.

Model Details

Model Description

Quant-Brain is a LoRA adapter fine-tuned on top of Upstage/SOLAR-10.7B-Instruct-v1.0 using QLoRA (4-bit NF4 quantization with double quantization). It is optimized for financial domain tasks including news sentiment classification, financial question answering, and providing strategic analysis for algorithmic trading systems.

The adapter targets both Attention layers (q/k/v/o projections) and MLP layers (gate/up/down projections) to enhance financial reasoning capability beyond simple pattern matching.

• Developed by: Quant-Brain Team
• Model type: Causal Language Model (LoRA Adapter)
• Language(s): English, Korean
• License: Apache 2.0
• Finetuned from: Upstage/SOLAR-10.7B-Instruct-v1.0

Model Sources

• Base Model: Upstage/SOLAR-10.7B-Instruct-v1.0
• Architecture: Llama-based with Depth Up-Scaling (DUS)

---

Uses

Direct Use

• Financial News Sentiment Analysis: Classify news articles as positive, negative, or neutral with reasoning.
• Financial Question Answering: Answer questions about investing, portfolio management, risk, and market mechanics.
• Trading Decision Support: Provide strategic analysis combining market context, ML predictions, and news sentiment for the Quant-Brain algorithmic trading pipeline.

Downstream Use

Designed as the "Brain" service in the Quant-Brain automated trading system:

1. Receives market context (OHLCV, technical indicators), ML predictions (XGBoost/LSTM), and news data.
2. Generates strategic decisions (BUY/SELL/HOLD) with confidence levels and reasoning.
3. Outputs structured JSON for downstream execution by the trading engine.

Out-of-Scope Use

• Not a financial advisor. This model is for research and educational purposes only.
• Not suitable for tasks requiring real-time, sub-second latency without GPU inference acceleration.
• Not designed for non-financial general-purpose chat or creative writing tasks.

---

Bias, Risks, and Limitations

• Financial Bias: Training data primarily covers US equities and English-language financial news; may not generalize to other markets or languages.
• Temporal Bias: Training data reflects market conditions up to 2024-2025; market regime changes may reduce accuracy.
• Sentiment Nuance: Neutral vs. mildly positive/negative boundary cases may be misclassified (observed in evaluation).
• Not Investment Advice: Outputs should never be used as sole basis for investment decisions.

Recommendations

• Always combine model outputs with human judgment and additional data sources.
• Monitor model performance over time for drift as market conditions change.
• Use ensemble approaches (multiple models) rather than relying on a single model.

---

How to Get Started with the Model

Quick Start (with PEFT)

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

BASE_MODEL = "Upstage/SOLAR-10.7B-Instruct-v1.0"
ADAPTER = "YOUR_HF_USERNAME/quant-brain-solar-10.7b-finance"

# Load with 4-bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

# Inference
prompt = """### User:
Analyze the sentiment of the following financial news. Classify as positive, negative, or neutral, and explain your reasoning.

NVIDIA's data center revenue surged 409% year-over-year, driven by explosive AI chip demand.

### Assistant:
"""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

vLLM Serving (Merged Model)

# First merge adapter with base model, then serve
python -m vllm.entrypoints.openai.api_server \
    --model /path/to/merged-model \
    --max-model-len 4096 \
    --dtype bfloat16 \
    --gpu-memory-utilization 0.90

---

Training Details

Training Data

A curated financial domain dataset of 147,324 training and 16,370 validation samples (163,694 total), sourced from:

All data formatted as instruction-input-output triplets in JSONL format.

Training Procedure

Preprocessing

• Instruction-tuning format: ### User:\n{instruction}\n{input}\n\n### Assistant:\n{output}
• Token packing enabled (median ~70 tokens/sample → ~7 samples per 512-token sequence)
• Padding to max sequence length of 512 tokens (P95 coverage)

Training Hyperparameters

Speeds, Sizes, Times

---

Evaluation

Metrics

Training Loss Curve

Step     Loss     Eval Loss    Token Accuracy
────     ────     ─────────    ──────────────
  10     2.903      —           52.8%
 100     1.132      —           68.9%
 500     0.830      —           79.4%
1000     0.795     0.807        80.0%
2000     0.768     0.777        80.5%
3000     0.755     0.769        80.3%
3376     0.772      —           80.1%

Functional Test Results

• Sentiment Accuracy: 3/4 (75.0%) — neutral/positive boundary case
• Financial QA Quality: All responses factually accurate and well-structured
• Average Latency: 5.6s (single A40, 4-bit quantized inference)

---

Environmental Impact

Carbon emissions estimated using the ML Impact Calculator.

• Hardware Type: NVIDIA A40 (48GB VRAM)
• Hours used: 22.7
• Cloud Provider: RunPod
• Compute Region: US
• Carbon Emitted: ~3.4 kg CO₂eq (estimated)

---

Technical Specifications

Model Architecture and Objective

• Base Architecture: Llama-2 based with Depth Up-Scaling (SOLAR DUS)
• Parameters: 10.7B base + ~134M trainable LoRA params (1.25% of total)
• Objective: Causal Language Modeling (next-token prediction) with SFT
• LoRA Strategy: Attention + MLP targeting for enhanced financial reasoning

Compute Infrastructure

Hardware

• GPU: NVIDIA A40 (48GB VRAM)
• System RAM: 100GB+
• Storage: NVMe SSD

Software

• PyTorch ≥ 2.1.0
• Transformers ≥ 4.40.0
• PEFT 0.18.1
• TRL ≥ 0.8.0
• BitsAndBytes ≥ 0.43.0
• Flash Attention 2 (SDPA fallback)

---

Citation

@misc{quant-brain-solar-finance-2025,
  title   = {Quant-Brain SOLAR-10.7B Financial Adapter},
  author  = {Quant-Brain Team},
  year    = {2025},
  url     = {https://github.com/YOUR_USERNAME/ai_quant_brain},
  note    = {QLoRA fine-tuned adapter for financial sentiment analysis and trading decisions}
}

Framework Versions

• PEFT 0.18.1
• Transformers 4.52.4
• TRL 0.18.1
• BitsAndBytes 0.46.0
• PyTorch 2.7.1+cu126