MPORTANT: when i say “which one would YOU prefer”, i mean this because im building this not only for myself.
There must exist people out there running into the same problem. If you are one of those, which one would make you smile?<\/p>\n
I’ve been building a community labeling platform for financial news sentiment \u2014 one label per asset, not generic.
The idea is that “OPEC increases production” is bearish for oil but FinBERT calls it bullish because it says something about “increasing” and “production.”
I needed Asset specific labels for my personal project and couldn’t find any, so i set out to build them and see who is interested.<\/p>\n
I now have ~46,000 labeled headlines across 27 securities (OIL, BTC, ETH, EURUSD, GOLD, etc.), generated by Claude Haiku with per-asset context.
Human validation is ongoing(only me so far, but i am recruiting friends). Im calling this v0.1.<\/p>\n
I want to train LoRA adapters on top of FinBERT, one per security, 4-class classification (bullish \/ bearish \/ neutral \/ irrelevant).<\/p>\n
Three paths I’m considering:<\/p>\n
- \n
- HuggingFace Spaces (free T4) Run training directly on HF infrastructure. Free, stays in the ecosystem. Never done it for training, only inference.<\/li>\n
- Spot GPU (~$3 total) Lambda Labs or Vast ai , SSH in, run the script, done in 30 min per adapter. Clean but requires spinning something up, will cost me some goldcoins.<\/li>\n
- Publish datasets only for now Or i could just push the JSONL files to HF as datasets, write model card stubs with “weights coming.” Labeling data is the hard part \u2014 training is mechanical. v0.1 = the data itself. But that is what i built it for, isnt it?<\/li>\n<\/ol>\n
My instinct is option 3 first, then spot GPU for the weights. But curious what people here would do \u2014 especially if you’ve trained on HF Spaces before.<\/p>\n
Project: <ask me> \u2014 contributions welcome if you want to label headlines.<\/p>\n
If you’re working on something similar, drop a comment \u2014 happy to share the export pipeline.<\/p>\n<\/div>\n
submitted by \/u\/Poli-Bert <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
