---
schema_version: "newruntime-agent-readable-v0.2"
type: "post"
stable_id: "post:fireworks-lora-fullft-three-test-protocol"
slug: "fireworks-lora-fullft-three-test-protocol"
title: "Run Three Tests Before Replacing LoRA With Full Fine-Tuning"
description: "Fireworks shows how data coverage, optimization, and adapter capacity can create or close an apparent quality gap between LoRA and full fine-tuning."
retrieval_nugget: "Fireworks shows how data coverage, optimization, and adapter capacity can create or close an apparent quality gap between LoRA and full fine-tuning. When a simple LoRA run loses to full fine-tuning, the adapter may not be the limiting factor. Fireworks tested Qwen3.5-9B on three programmatically scored tasks and organized the diagnosis around data coverage, optimization, and adapter capacity."
status: "published"
published_at: "2026-08-03"
updated_at: "2026-08-03"
record_date: "2026-08-03"
date_kind: "published_at"
topics: ["post-training","open-models","evals","inference"]
source_urls: ["https://fireworks.ai/blog/three-tests-to-run-before-you-switch-from-LoRa-to-FullFT"]
visuals: [{"id":"fireworks-lora-fullft-three-test-protocol","kind":"editorial-diagram","role":"hero","src":"https://newruntime.com/images/posts/fireworks-lora-fullft-three-test-protocol.webp","alt":"Hand-drawn diagnostic ladder that tests broader training coverage, a tuned optimization recipe, and increased adapter rank before comparing LoRA and full fine-tuning on quality and total deployment cost.","caption":"A FullFT advantage becomes meaningful only after coverage, recipe, and adapter-capacity alternatives have been tested under a declared comparison budget.","credit":"New Runtime synthesis from Fireworks AI","source_url":"https://fireworks.ai/blog/three-tests-to-run-before-you-switch-from-LoRa-to-FullFT","generated_with":"gemini-3.1-flash-image","width":1600,"height":900,"legend":[{"label":"Coverage","description":"Broaden task variation before interpreting an out-of-distribution failure as a method limit."},{"label":"Optimization","description":"Sweep learning rate and schedule because the best adapter recipe can differ sharply from FullFT."},{"label":"Capacity","description":"Increase rank under controlled scaling and check whether validity and exact quality move together."},{"label":"Economics","description":"Compare training and serving costs after both methods meet the required quality bar."}]}]
routes: {"html":"https://newruntime.com/posts/fireworks-lora-fullft-three-test-protocol/","markdown":"https://newruntime.com/posts/fireworks-lora-fullft-three-test-protocol.md","json":"https://newruntime.com/posts/fireworks-lora-fullft-three-test-protocol.json"}
source_format: "markdown"
---

# Run Three Tests Before Replacing LoRA With Full Fine-Tuning

## Retrieval answer

Fireworks shows how data coverage, optimization, and adapter capacity can create or close an apparent quality gap between LoRA and full fine-tuning. When a simple LoRA run loses to full fine-tuning, the adapter may not be the limiting factor. Fireworks tested Qwen3.5-9B on three programmatically scored tasks and organized the diagnosis around data coverage, optimization, and adapter capacity.

When a simple LoRA run loses to full fine-tuning, the adapter may not be the limiting factor. Fireworks tested Qwen3.5-9B on three programmatically scored tasks and organized the diagnosis around data coverage, optimization, and adapter capacity.

Coverage moved the placement task first. Increasing the dataset tenfold raised LoRA validity from 34.33% to 68.33%. Tuning the learning rate then brought rank-32 LoRA to 82.67% on a common suite against 83% for FullFT, a 0.33-point difference inside the reported confidence interval. On an untouched final suite, the selected LoRA recipe scored 81.67% versus 80.33% for FullFT.

Capacity mattered on register allocation. Raising rank from 8 to 128 increased valid allocations from 80.5% to 90.5%, while FullFT reached 93%. The stricter exact-decision metric still favored FullFT, showing why validity and imitation quality should be measured separately.

The Nexa virtual-machine task exposed both recipe and curriculum effects. A learning-rate sweep reversed an apparent 31-point FullFT advantage on supported program lengths. Beyond the training horizon, randomizing register starts raised exact trace accuracy to 100% for LoRA and 99.22% for FullFT, demonstrating that rank cannot replace examples of unseen positions.

The comparison changed again when three tasks shared one fixed training budget. Tuned rank-128 LoRA reached 76.60%, while FullFT reached 80.89%. FullFT used 8.6% more active training compute and retained a 4.29-point lead. When per-task exposure was preserved instead of diluted, LoRA led the tested FullFT recipe.

For New Runtime, the decision should be an experiment protocol rather than a method preference. Broaden support, tune each method independently, test adapter rank, declare what the budget holds constant, and then include serving economics: many LoRAs can share one base deployment, while every FullFT variant needs dedicated capacity.