Llama 3.2 1B, deployed on Android with 480 examples

Llama 3.2 1B was LoRA fine-tuned on 480 hand-curated ShareGPT examples, then merged and quantized to Q4_K_M before being shipped in an Android app.

• Base model: meta-llama/Llama-3.2-1B-Instruct
• Tuning: LoRA r=16, alpha=32, 7 target modules, 3 epochs, lr=2e-4
• Final size: 770MB
• Runtime: llamadart with the llama.cpp GPU backend on Android

On-device measurements showed the following performance.

• Pixel 7 Pro: about 25 tok/s, 1.5s load, 3-4s for 80-token generation
• Pixel 6a: about 15 tok/s, 2.5s load, 4-5s for 80-token generation
• Samsung S22: about 22 tok/s, 2.0s load, 3.5-4.5s for 80-token generation

In evaluation on 48 held-out samples across 4 criteria — on-tone, specific, funny, and category-correct — it achieved an 80%+ pass rate.

The most notable point was the data scale. At the 480 level, performance had already plateaued from around 400 examples, and the bottleneck was data quality rather than the number of examples.