FinGPT - AI Language Models Tool

Overview

FinGPT is an open-source project that provides domain-adapted large language models, training scripts, instruction‑tuning datasets, and end-to-end pipelines to build FinLLMs (financial LLMs). The project emphasizes data-centric tooling (an automated pipeline that ingests ~30+ web sources), lightweight adaptation (LoRA / QLoRA / 8-bit workflows) and retrieval-augmented methods so researchers and practitioners can create, evaluate, and update finance-focused LLMs without the cost of full re‑pretraining. According to the repository and associated papers, FinGPT publishes multi-task and single-task models (examples: sentiment, relation extraction, forecaster) and provides pretrained LoRA adapters on Hugging Face alongside reproducible notebooks and demo Spaces. ([github.com](https://github.com/AI4Finance-Foundation/FinGPT)) The project also maintains an instruction-tuning benchmark and multiple datasets (sentiment, headline classification, FiQA Q&A, forecaster time-series dataset) to measure FinLLM capabilities and compare base models (Llama‑2, Falcon, MPT, ChatGLM2). FinGPT includes an example “FinGPT-Forecaster” Hugging Face Space that demonstrates a retrieval + LLM forecasting pipeline and released dataset splits for Dow30 and SZ50 forecaster tasks. Community traction is high (GitHub stars/forks and active Hugging Face models/datasets), and the authors have published several arXiv/Workshop papers describing the architecture and benchmarks. ([huggingface.co](https://huggingface.co/FinGPT))

GitHub Statistics

• Stars: 18,407
• Forks: 2,589
• Contributors: 29
• License: MIT
• Primary Language: Jupyter Notebook
• Last Updated: 2025-12-06T08:50:27Z

Key Features

• Pretrained FinLLM LoRA adapters for sentiment, relation extraction, forecasting tasks.
• Data pipeline that curates time‑sensitive financial sources (30+ web sources) for continual updates.
• Instruction‑tuning benchmark suite (FinGPT-Benchmark) covering sentiment, NER, headline and QA.
• Resource‑efficient fine‑tuning recipes: LoRA, QLoRA, 8‑bit training, runnable on consumer GPUs.
• Retrieval‑augmented (RAG) forecaster demo and public forecaster datasets for Dow30 and SZ50.

Example Usage

Example (python):

## Minimal example: load a FinGPT LoRA adapter from Hugging Face and run a short inference
# Model example: "FinGPT/fingpt-mt_llama2-7b_lora" (see Hugging Face FinGPT org)
# Requires: transformers, accelerate, peft (pip install transformers accelerate peft)
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# 1) load tokenizer for the LoRA adapter (adapter repo usually bundles tokenizer info)
adapter_repo = "FinGPT/fingpt-mt_llama2-7b_lora"
# 2) choose a compatible base model (example shown generically)
base_model = "meta-llama/Llama-2-7b-chat-hf"  # replace with the exact base used by the adapter if required

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(adapter_repo, use_fast=False)

# Load base model then apply LoRA adapter with PEFT
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto", torch_dtype=torch.float16)
model = PeftModel.from_pretrained(model, adapter_repo, device_map="auto")
model.eval()

# Inference
prompt = "Summarize the recent news sentiment for AAPL in one paragraph."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.2)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# Notes: replace base_model with the exact base checkpoint used by the adapter on Hugging Face.
# See the FinGPT GitHub README and Hugging Face model pages for model-specific instructions and compatibility.

Benchmarks

FinGPT v3.3 - Weighted F1 (sentiment benchmark): 0.882 (Source: https://github.com/AI4Finance-Foundation/FinGPT#benchmark-results)

FinGPT v3.3 - FiQA-SA score (sentiment subtask): 0.903 (Source: https://github.com/AI4Finance-Foundation/FinGPT#benchmark-results)

OpenAI Fine-tune - Weighted F1 (comparison): 0.878 (Source: https://github.com/AI4Finance-Foundation/FinGPT#benchmark-results)

FinGPT (QLoRA) - Weighted F1 (quantized LoRA): 0.777 (Source: https://github.com/AI4Finance-Foundation/FinGPT#benchmark-results)

Base Llama2-7B - Weighted F1 (baseline): 0.390 (Source: https://github.com/AI4Finance-Foundation/FinGPT#benchmark-results)

Key Information

• Category: Language Models
• Type: AI Language Models Tool