FinLoRA — SEC 10-K Financial Document QA
Phi-3.5-mini-instruct (3.8B) fine-tuned on 34,799 SEC 10-K examples via QLoRA supervised fine-tuning.
Trained to answer financial questions grounded in document context, with mandatory source citations.
| Metric | Base | Fine-tuned |
|---|---|---|
| ROUGE-1 | 0.34 | 0.64 |
| BERTScore F1 | 0.86 | 0.93 |
| Numerical EM | 35.8% | 52.2% |
GitHub · SFT adapter · DPO adapter
Running on CPU — output streams token by token (~90–120s total). Paste any 10-K passage and ask a question.
Example questions
Model: Phi-3.5-mini-instruct + QLoRA SFT on 34,799 SEC 10-K examples. Aayush Patel · Georgia Tech