Member-only story
An approachable, reproducible guide to LoRA, quantization, training loss, validation, and model evaluation.
When I started this project, I wanted more than a better score. I wanted to
understand what fine-tuning was actually doing. This guide starts with the basic ideas, explains the vocabulary as it comes up, and uses our real training runs to show what worked and what failed.
*If you are not a paid medium member, u can read it *here
The result:after six controlled experiments, validation Macro-F1 rose
from0.2956 to 0.8246. Exact-match accuracy reached91.5%, and every
response in the 200-example validation set could be parsed.
This work was built as part of NVIDIA’s India Agentic AI Open Hackathon, where our team, RTX Renegades , built NemoCounsel, an AI contract review assistant powered by the fine-tuned model described in this post. Credit to my teammates Aditya Rallapalli,
[,](https://www.linkedin.com/in/sainathch45/)
**Sainath Chakravadhanula**[, and](https://www.linkedin.com/in/ruchita-b-424893256/)
Ruchita Bhandarifor building this with me, and to our NVIDIA mentors
Stuti Chourasia
Raghul Asokan****and
Sutanshu Raj****for their guidance throughout. Thank you to
NVIDIAfor the GPU compute support that made these training runs possible.