AppGen Qwen3-VL frozen SFT — A-variant arm G
This is arm amex_nav_lr2p5e7 from a preregistered four-arm follow-up to the
best NGC configuration A. The language model was fully fine-tuned while the
complete Qwen3-VL visual tower, merger, and deep-stack mergers remained
frozen. Publication verifies every model.visual tensor byte-for-byte against
the pinned base and verifies that language-model weights changed.
Training contract
- Base: Qwen/Qwen3-VL-8B-Instruct at commit 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
- Sources: luca0621/appgen-sft-ngc-v1 at 769ea99dbc4ff190048ae0db37eb6310dba595e0 and
Yuxiang007/AMEX at 17196b29c88dd48a7fb90ef9131bc5c7bf39f26e
- Dataset variant: amex_nav
- Dataset SHA-256: 8c15fba4551322e8eafdaf9a733d9ef1076764231f302514efd79187cd553669
- Exposures: 3,588; unique semantic examples: 3,588
- Direct-grounding exposures: 400
- Completion-retention replay exposures: 420
- Image provenance: AMEX real-device navigation screenshots plus synthetic AppGen grounding/completion screenshots
- Coordinates: normalized 0–1000
- Prompt: proven A prompt; SHA-256 67ff8adb0e78a617f3d0edcf196d4e4cc3239967a8c30619fc5afc14484ee8c0
- Optimizer: full-language AdamW, learning rate 2.5e-07, cosine schedule, 5% warmup
- Batch: microbatch 4 × two GPUs × gradient accumulation 4 = global batch 32
- Epochs: 1; optimizer updates: 113
- Loss scale: Swift default
- Frozen: visual encoder and aligner; trainable: language model and LM head
Only the final checkpoint (checkpoint-113) is published.
The exact system prompt is appgen_system_prompt.txt; run_manifest.json records
exact hashes, the dataset receipt, overlap audit, and weight verification.
Intermediate 25-step checkpoints remain local.
Evaluation and limitations
The training/evaluation overlap gate compares EXIF-transposed decoded RGBA
pixels plus image dimensions against the seven pinned AW7 test suites. It also
compares normalized instruction hashes. The exact pinned receipt must report
zero train/eval pixel and instruction overlaps for every data variant and suite
before publication. Static grounding benchmarks are not a substitute for
interactive AndroidWorld task success, and no score is claimed in this card.
This model is for Android visual-agent research, not safety-critical autonomous
deployment.
