fatsam13's picture
Publish v2 single food and nutrition vision adapter
ad41439 verified
|
Raw History Blame Contribute Delete
934 Bytes

Version 2 Release Notes

Version 2 is one multi-task LoRA adapter, distributed as matching PEFT and GGUF representations. It replaces the current adapter for new integrations while the v1 GGUF remains available for rollback.

Changes

  • Improves Food-101 strict accuracy from 86.80% to 87.60% on 500 held-out images.
  • Adds experimental visible-ingredient and rough-gram JSON output.
  • Keeps nutrition calculation outside the model through USDA FoodData Central mappings and application rules.
  • Preserves rank 8, alpha 16, and the seven original target-module families.
  • Validated with the existing Unsloth UD-Q4_K_XL base and mmproj-F16.gguf.

iOS Migration

Update only the adapter download filename to:

beforeat-food-nutrition-vision-lora-v2-f16.gguf

Continue to reuse the existing base GGUF and projector, and apply the adapter at scale 1.0. Do not load the historical v1 adapter at the same time.