WORK IN PROGRESS – EXPERIMENTAL RESEARCH PROJECT
A continuously learning transformer language model that trains automatically every ~1.5 hours on diverse datasets using GitHub Actions.
Important Notice
Fin.AI is an experimental research prototype and work in progress.
The model is under continuous training and may produce inaccurate, inappropriate, biased, or nonsensical outputs.
Do NOT use for production applications, critical systems, or high-stakes decisions.
Use at your own risk.
Fin.AI is an experimental GPT-style language model that trains 24/7 with a rotating curriculum of 24 different dataset families.
Core characteristics:
- Fully automated hourly training (GitHub Actions)
- 24 diverse dataset categories (news, math, code, dialogue, science, instructions...)
- Focus rotates every ~1.5 hours → targeted capability improvement
- Models automatically pushed to Hugging Face after each run
- Training metrics publicly visible on Weights & Biases
- Designed to run efficiently even on free GitHub runners
This is not a production-ready model. Expect evolving (and sometimes unstable) behavior.
| Feature | Description |
|---|---|
| Automated Continuous Training | Trains every ~1.5 hours – completely hands-free |
| Rotating Curriculum | 24 dataset families covering very different capabilities |
| Hugging Face Integration | Latest checkpoint pushed automatically after every training cycle |
| Real-time Monitoring | Full metrics, loss curves and samples on Weights & Biases |
| Flexible Scale | Easily switch between ~15M and ~350M+ parameters |
| CPU-friendly | Optimized to train efficiently on standard GitHub Actions runners |
| Category | Example Datasets | Cycle Hours | Primary Focus |
|---|---|---|---|
| Encyclopedia | WikiText | 0, 6 | Broad world knowledge |
| Creative Writing | TinyStories | 1, 18 | Storytelling & narrative |
| News | CNN, AG News, CC News | 2,15,17,20 | Current events & factual style |
| Math & Reasoning | GSM8K, CommonsenseQA | 3,9,19,23 | Problem solving & logic |
| Open Web Text | OpenWebText, C4 | 4,11 | Diverse internet language |
| Question Answering | SQuAD | 5,22 | Reading comprehension |
| Instruction Following | Alpaca, Dolly | 7,14,21 | Following user instructions |
| Reviews & Sentiment | IMDB, Amazon, Yelp | 8,10,16 | Opinion & sentiment analysis |
| Scientific / Medical | PubMed | 12 | Scientific & medical literature |
| Conversations | UltraChat | 13 | Natural dialogue |
from huggingface_hub import hf_hub_download
hf_hub_download("MeridianAlgo/Fin.AI", "model.pt", local_dir="./model")
hf_hub_download("MeridianAlgo/Fin.AI", "config.json", local_dir="./model")from fin_ai.model import FinAIModel
import torch
model = FinAIModel.from_pretrained("./model")
tokenizer = model.tokenizer
prompt = "The future of artificial intelligence is"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100, temperature=0.8)
print(tokenizer.decode(outputs[0]))Warning: Output quality is experimental and may contain factual errors, biases, or inappropriate content.
| Preset | Parameters | Layers | Heads | Hidden Dim | Recommended Use Case |
|---|---|---|---|---|---|
| tiny | ~15M | 6 | 4 | 256 | Very fast experiments |
| small | ~40M | 8 | 8 | 512 | Default – good CPU performance |
| medium | ~120M | 12 | 12 | 768 | Noticeably higher quality |
| large | ~350M | 24 | 16 | 1024 | Best results (GPU recommended) |
- Latest checkpoint: huggingface.co/MeridianAlgo/Fin.AI
- Training pipeline runs: GitHub Actions
- Live metrics & samples: Weights & Biases
Made with passion by the Fin.AI team
⭐ Star on GitHub 🤗 View & download on Hugging Face