GitHub Repos advanced 3 min read May 7, 2026
Public Preview Sign in free for the full digest →

FinGPT: Financial GPT for researchers and startups

“You can fine-tune a financial LLM that beats GPT-4 on sentiment for $17 — Bloomberg spent $3 million on theirs.”

FinGPT: Financial GPT for researchers and startups
1 Views
0 Likes
0 Bookmarks
Source · github.com

“"The total cost of fine-tuning FinGPT is less than $300 (training from scratch is expensive, e.g., BloombergGPT costs about $3M)" — FinGPT README, AI4Finance Foundation (https://github.com/AI4Finance-Foundation/FinGPT)”

You want to add financial NLP capabilities — sentiment analysis, headline classification, named entity recognition — to your application, but the options are either a $3M proprietary model (BloombergGPT) or a general-purpose LLM (GPT-4) that was never explicitly trained on financial language. Fine-tuning your own domain-specific model seems expensive and technically daunting: you need a labeled dataset, a training pipeline, a GPU budget, and a way to keep the model current as financial language evolves. RLHF-style training requires human annotators for every new finance domain you target.

llmfinancefine-tuningopen-sourcepythonpytorchrlhf

FinGPT takes an existing open-source base model (say, LLaMA 2 7B) and fine-tunes it on financial text using LoRA — a technique that freezes 99.94% of the model's weights and only trains a tiny 3.67M-parameter adapter layer, which is why the whole process costs $17 instead of $3M. The training signal for sentiment comes from RLSP: instead of paying humans to label headlines as positive or negative, the framework automatically labels each news item based on the stock price movement that followed the announcement — price went up means the news was positive. You feed this self-labeled dataset of 620K+ headlines into the LoRA fine-tuner for 3 epochs, and what comes out is a LoRA adapter that plugs on top of your base model. At inference time, you load the base model plus the adapter, and it handles financial text classification tasks well.

01
RLSP training signal — uses post-announcement stock price movements as free automatic labels for 620K+ headlines, so you skip expensive human annotation entirely
02
LoRA fine-tuning for $17–$300 — reduces trainable parameters from 6.17B to 3.67M, making a full fine-tuning run feasible on a single RTX 3090 in under 18 hours
03
Pre-trained LoRA adapters on HuggingFace — you can skip training entirely and load fingpt-mt_llama3-8b_lora or the Llama2-13B sentiment adapter directly
04
Six instruction-tuning datasets — covers sentiment (76.8K rows), relation extraction (27.6K), headline classification (82.2K), NER (511), financial QA (17.1K), and Chinese finance exam (1.1K)
05
Multi-base-model support — same training recipes work across LLaMA 2 7B/13B, Falcon, MPT, BLOOM, ChatGLM2, Qwen, InternLM, and LLaMA 3
06
Five-layer data pipeline — aggregates from 34+ sources including CNBC, Reuters, Yahoo Finance, and SEC filings with real-time updates
07
FinGPT-Forecaster demo — DOW30 stock prediction app on HuggingFace Spaces, runnable without any local setup
Who it’s for

If you are an ML engineer or researcher who needs a reproducible, low-cost baseline for financial NLP classification tasks — sentiment analysis, headline categorization, named entity recognition — FinGPT gives you a full fine-tuning pipeline and pre-trained adapters to build on. It is also well-suited if you are studying RLHF alternatives and want to evaluate RLSP as a human-annotation-free labeling technique. Not useful if you need financial question answering, document summarization, or stock movement prediction — independent benchmarks show near-random performance on those tasks, and the p...

Worth exploring

Worth exploring if your specific use case is financial sentiment classification or headline categorization and you want a reproducible, MIT-licensed baseline that is cheap to retrain on new data. The RLSP methodology is genuinely clever and academically validated (Best Presentation at IJCAI 2023, two NeurIPS 2023 Workshop papers). However, if you need financial QA, reasoning over documents, or summarization, skip it entirely — independent benchmarks show 3.8–28% exact match vs. GPT-4's 69–76%, and ROUGE-1 near zero on summarization. By April 2026 even the sentiment classification edge over frontier models has narrowed significantly.

Developer playbook
Tech stack, code snippet, sentiment, alternatives.
PM playbook
Adoption angles, user fit, positioning.
CEO playbook
Traction signals, ROI, build vs buy.
Deep-dive insight
Full long-form analysis, no fluff.
Easy mode
Core idea, fast — when you need the gist.
Pro mode
Technical nuance, edge cases, tradeoffs.
Read the full digest
Go beyond the preview

Deep-dive insight, Easy and Pro modes, plus action playbooks — the full breakdown is one tap away.

Underrated tools. Unfiltered takes.

Read the full digest in the Snaplyze app for deep-dive insight, Easy and Pro modes, and the playbooks you can actually use.

Install Snaplyze →