Use case · Fine-tuning (SFT + LoRA)
Content safety
A child-facing assistant has to answer at the right level for each age group, and stay appropriate in tone, without heavy hardware.
The experiment
Fine-tune TinyLlama-1.1B to give age-appropriate educational answers across developmental stages, then find the LoRA setup with the best balance of quality and cost.
- Dataset
- children_sft_dataset, expanded to 1,200 samples across 5 age groups and 4 intent types.
- Agent
- TinyLlama-1.1B fine-tuned with SFT and LoRA on a T4 GPU, varying learning rate, target modules, and LoRA rank.
- Knobs tested
- Learning rateLoRA target modulesLoRA rank
What the experiment showed
+16.9%
ROUGE-L over baseline for the winning setup
+6.2%
token accuracy over baseline
0.3%
extra runtime for those gains
- 01Expanding LoRA to all linear modules at rank 32 gave the best results for the cost.
- 02Raising the learning rate made training unstable and lowered BERT-F1.
- 03Combining every knob scored highest on raw metrics, but at a runtime cost that made it the wrong pick.
- 04Age fit is hard to measure: overlap metrics miss it on open-ended tasks, and tone can slip for older age groups.
How RapidFire helped
Every configuration in one run
Compared the knobs side by side on the same hardware, instead of one run after another.
Live results
Scored each version on slices of the eval set as it ran, so decisions came early.
Weak runs stopped early
Cut directions that were not working and put the compute behind the ones that were.
Winners cloned
Cloned the strongest run and changed one factor at a time to find what drove the gain.
Apply it to your data
- 1Label your examples by audience and intent.
- 2Sweep module coverage, rank, and learning rate on modest hardware.
- 3Weigh quality against runtime, and review outputs for tone by audience.