Gemma-3 Finance Mix
A lightweight Gemma-3 270M model fine-tuned for financial Q&A, causal-lm/finance, news-headline sentiment and retail-investor discourse.
Overview
| Item | Details |
|---|---|
| Base checkpoint | google/gemma-3-270m-it |
| Fine-tune method | LoRA (PEFT) with Unsloth |
| Training run | 1 epoch • 325,528 blended examples • 100 steps |
| Trainable params | 30.4 M / 298 M (10.18 %) |
| Loss | 4.11 → 2.74 |
| Hardware | 2 × T4-16GB (Collab Free Tier) |
| License | Apache-2.0 |
| Intended use | Educational & research |
Datasets
| Dataset | Size | Focus |
|---|---|---|
gbharti/finance-alpaca |
52 k | Instruction Q-A on corporate finance & investing |
Balaji173/finance_news_sentiment |
217 k | Bullish/bearish labels for news headlines |
winddude/reddit_finance_43_250k |
250 k | Reddit finance post–comment pairs |
causal-lm/finance |
31 k | Analytical prompts & causal reasoning in economics/markets |
All shards were concatenated and wrapped with the Gemma chat template before training.
Responsible use
Disclose AI assistance, double-check outputs, and do not rely on this model for trading decisions. The author and base-model creators accept no liability for financial losses.