LoRA vs Full Fine-Tuning — Flappy (fixed_l2) Ablation
Evaluation results for an ablation comparing LoRA vs full-parameter fine-tuning of
StarVLA/Qwen3-VL-4B-Instruct-Action (openvla framework) on the Flappy environment at a
fixed latency of 2 raw frames (66.67 ms).
Result (latency 2, 50 episodes, seed 42)
| Method | Mean reward | Std | Mean survival steps | Final eval CE loss | Model |
|---|---|---|---|---|---|
| Full FT | 18.52 | 18.93 | 167.6 | 0.060 | openvla_flappy_l2_full |
| LoRA (r=32, α=64) | 3.86 | 0.27 | 49.6 | 0.224 | openvla_flappy_l2_lora |
Full fine-tuning outperforms LoRA by ~4.8× in reward and survives ~3.4× longer. On this imitation task (CE on a Sample-Factory teacher), LoRA's low-rank update underfits relative to full FT, consistent with the higher held-out CE loss.
Controlled setup (everything identical except the variable under test)
- Data:
latency-sensitive-bench/flappy_fixed_l2_fs4(990 episodes, 948,645 frames), frame stack 4. - Loss: cross-entropy (categorical 2-action head). Steps: 2000. Seed: 42.
- Effective (mini) batch = 64 on both sides, tuned per GPU to fill memory:
- Full FT: per_device 32 × grad_accum 2 (single B200, ~151 GB peak).
- LoRA: per_device 16 × grad_accum 4 (single A100-80GB, ~73 GB peak; 32 OOMs).
- Learning rate (the only deliberately method-specific knob): VLM backbone LR 2e-5 (full) vs 2e-4 (LoRA, ~10×); the from-scratch action head is 1e-4 for both. Rationale: full FT updates all pretrained weights (small LR avoids catastrophic forgetting); LoRA's effective update is scaled by α/r and needs a ~10× larger LR — a shared LR would under-tune LoRA and bias the comparison.
- Eval:
latency_bench.run,fixedlatency = 2, 50 episodes. Both models were evaluated on the same A100 hardware for an identical eval environment (reward is deterministic given seed).
Files
full/queue_eval_latency_2.json,full/episode_metrics.jsonllora/queue_eval_latency_2.json,lora/episode_metrics.jsonl