// the find
AI4Finance-Foundation/FinGPT
FinGPT: Open-Source Financial Large Language Models! Revolutionize 🔥 We release the trained model on HuggingFace.
FinGPT is an umbrella research project from AI4Finance Foundation bundling LoRA-finetuned financial sentiment models, a Dow-30 stock-movement forecaster, and RAG/benchmark pipelines on top of open base models (Llama2, Falcon, ChatGLM2, etc.). It's for researchers and quant-adjacent developers who want a cheap, reproducible alternative to BloombergGPT, not a plug-and-play trading product.
The cost/benchmark transparency is unusually good — they show fine-tuning a 13B sentiment model on a single RTX 3090 for $17.25, versus BloombergGPT's $2.67M on 512 A100s, with actual F1 numbers on FPB/FiQA/TFNS/NWGI to back it up. Multiple pre-trained LoRA adapters (Llama2, Falcon, MPT, Bloom, ChatGLM2, Qwen, InternLM) are already published on HuggingFace, so you're not starting from zero. There's real peer-reviewed backing — accepted papers at NeurIPS workshops, ICAIF, and IJCAI, not just blog-post claims. FinGPT-Forecaster ships a working HuggingFace Space demo and a Dockerfile, so you can poke at the stock-prediction use case without owning a GPU.
The repo is a sprawl rather than a coherent library — FinGPT_Others alone has parallel v1/v2 trading and robo-advisor folders, with checked-in nohup.out logs, pickled dataframes, and loose CSVs sitting next to the code. Most of the actual work lives in Jupyter notebooks, not the installable package; `pip install -e .` exists but doesn't get you much. The Forecaster is trained only on Dow 30 US data — anything else means collecting your own dataset and retraining, with no generalized data pipeline provided. Local inference of the models worth using needs real GPU hardware (RTX 3090 minimum), and the cloud fallback defaults to GPT-3.5-turbo, which is a strange choice to still be shipping as the default in 2026.