Instructions to use gavinlaw/qwen3vl4b-androidworld-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use gavinlaw/qwen3vl4b-androidworld-lora with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
LoRA adapters and frozen-checkpoint diagnostics for audited RL on AndroidWorld
Policy adapters (rank 64, alpha 128) for Qwen/Qwen3-VL-4B-Instruct trained online on AndroidWorld, judge adapters for Qwen/Qwen3-VL-2B-Instruct fine-tuned on trusted labels, and the per-trajectory gradient statistics of fully audited collections at frozen policy checkpoints. Adapters are PEFT directories (adapter_config.json + fp32 adapter_model.safetensors); load one with peft.PeftModel.from_pretrained(base, "<dir>"). Rollouts and screenshots are in the dataset androidworld-qwen3vl4b-rollouts of the same owner.
| path | contents | produced by |
|---|---|---|
runs/20261024/state-seg{1,4,5,6,7}.tar |
four-arm online pilot (full audit, judge only, learned and uniform 5% audits): train-pilot-*/adapter-rNNNN/, state.pt (resume state) and per-round logs, at the end of each run segment |
code/serve/pilot_arms.sh, code/scripts/upload_state.py |
runs/20261026/pilot2-state.tar |
the same for the confirmatory pilot with a shared KL step controller | same |
runs/20261026/pc1-state.tar |
full-audit positive-control run; adapter-r0012 is the drifted checkpoint theta_12 |
code/scripts/train_online.py --arm full |
runs/20261026/frozen/ |
fully audited collection at theta_0 (metadata, per-step features, step-gradient Gram) | code/serve/frozen_chain.sh, code/scripts/frozen_diagnostic.py |
runs/20261027/b-confirm/ |
fully audited collections at theta_0 and theta_12 (528 rollouts each) with Grams and metadata | code/serve/fresh_chain.sh |
runs/20261027/owl0/, runs/20261027/owl0-diag/ |
recorded policy calls, metadata and Gram of a collection with mPLUG/GUI-Owl-1.5-4B-Instruct |
code/serve/frozen_chain.sh |
runs/20261027/sft-out.tar |
judge adapters of Qwen3-VL-2B fine-tuned on n = 250-4000 trusted labels (three epochs each) and their development scores | code/scripts/judge_sft.py, code/scripts/judge_score.py |
runs/20261027/d25-fresh/ |
first pre-registered fresh confirmation: Grams, metadata, recorded calls, judge scores | code/serve/fresh_chain.sh, code/scripts/upload_run_outputs.py |
runs/20261028/d35-fresh/ |
second pre-registered fresh confirmation (704 rollouts per checkpoint): Grams, metadata, recorded calls, scores of five judges | same |
runs/20261028/judge-recipe-sweep/out/ |
judge adapters of six one-factor recipe changes and their scores | code/scripts/b_recipe_sweep.py |
Every path is pinned by the commit that added it; the paper lists the commits it uses. Sources and licenses: AndroidWorld (Apache-2.0), Qwen3-VL models (Apache-2.0), GUI-Owl (see its model card); the apps shown are the open-source apps bundled with AndroidWorld.
- Downloads last month
- -
Model tree for gavinlaw/qwen3vl4b-androidworld-lora
Base model
Qwen/Qwen3-VL-2B-Instruct