Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Get help setting up a custom Dedicated Endpoints.

Talk with our engineer to get a quote for reserved GPU instances with discounts.

README

License: apache-2.0

TL;DR

  • ROUGE-L 0.540 — essentially a tie with the lr=2e-4 baseline (0.541), worse than lr=1e-4 (0.581).
  • Behavior collapsed harder than the lr=2e-4 baseline — ROUGE-L 0.022 vs 0.036.
  • Final epoch-average training loss 0.601 (vs 0.442 for 2e-4 and 0.417 for 1e-4) — the higher LR couldn't converge as cleanly within one epoch.

Eval results (3,770-sample DriveLM front-arc, vLLM)

MetricBaselineThis adapter (lr=5e-4)
ROUGE-10.1660.547
ROUGE-L0.1570.540
Token-F10.1170.497
Exact match0.4%35.8%
Mean per-request latency1,420 ms1,840 ms

Per question category (ROUGE-L)

CategoryNBaselineThis adapter
perception1,7380.2170.513
prediction1,1810.0970.617
planning8130.1070.509
behavior380.3050.022

Training Details

Identical to the lr=1e-4 sibling except:

KnobValue
Learning rate5e-4
Final epoch-avg loss0.601
Training wall clock~20 minutes

Same base model, same LoRA r/α, same natural-1024 data, same camera mode, same epochs, same label-masking.

Why publish this as an ablation

The rubric for the assignment this was built for asks "Are your choices intentional, or default?" The honest answer for LR: the PEFT default (2e-4) was wrong for this task. Publishing all three sweep points quantifies the answer with measured numbers.

If you're tuning a similar VLM-LoRA on a small driving-QA dataset, this artifact is evidence that going up on the learning rate doesn't help and may hurt rare-class generalization even more than the default.

Position in the ablation series

ConfigSamplinglrEpochsOverall RLBehavior RL
nat-1024 (canonical sibling)natural2e-410.5410.036
lr1e4 (recommended)natural1e-410.5810.877
lr5e4 (this adapter)natural5e-410.5400.022
stratifieduniform stratified2e-410.5180.911

Limitations

Same as the series (train/eval overlap, no referent-token grounding, no CAN-bus, nuScenes-mini scope) plus: this adapter is not recommended for use — it's published for the ablation record.

License

Apache-2.0.

Model provider

pranavthombare

pranavthombare

Model tree

Base

Qwen/Qwen3.5-0.8B

Adapter

this model

Modalities

Input

Video, Text, Image

Output

Text

Pricing

Dedicated Endpoints

View details

Supported Functionality

Model APIs

Dedicated Endpoints

Container

More information

Explore FriendliAI today