What it does

FinGPT is a collection of open-source financial LLMs built by fine-tuning general-purpose models like Llama2 and ChatGLM2 on domain-specific datasets. The project covers sentiment analysis, headline classification, relation extraction, named-entity recognition, and a robo-advisor forecaster that predicts stock price movement from news and financials. Everything is published on HuggingFace, with training notebooks you can run on a single RTX 3090.

The interesting bit

The project’s core argument is economic, not architectural. BloombergGPT reportedly cost ~$3M and 53 days to train; FinGPT’s LoRA fine-tuning runs under $300 and updates weekly. The README is essentially a cost-accounting takedown dressed up as research — complete with GPU-hour spreadsheets comparing their $17.25 sentiment run against Bloomberg’s estimated $2.67M. Whether the comparison is fair (different scopes, different base models) is left as an exercise for the reader.

Key highlights

• FinGPT v3.3 (Llama2-13B + LoRA) claims weighted F1 of 0.882 on financial sentiment, trained in 17 hours on one RTX 3090
• FinGPT-Forecaster demo takes a ticker, date, and news window, then outputs analysis plus next-week price prediction
• Six instruction-tuning datasets published on HuggingFace, from 511-row NER to 82K-row headline tasks
• RLHF is flagged as the “secret ingredient” missing from BloombergGPT, though implementation details are sparse
• Multiple papers accepted at NeurIPS 2023 (Instruction Workshop) and ICAIF-23

Caveats

• The “better than GPT-4” claim applies only to fine-tuned sentiment on specific benchmarks, not general capability
• README mixes aspirational roadmap items with shipped features; some sections read like grant proposals
• RTX 3090 cost assumptions ($1/hour) rely on third-party GPU rental platforms, not stable cloud pricing

Verdict

Worth exploring if you’re building financial NLP on a budget or need a sentiment baseline that beats generic models. Skip if you need production-grade reliability guarantees or audited trading logic — this is research tooling with ambitious marketing.