appgen-qwen3-uedgrpo-simple-path-dense-s100-step100

AppGen Qwen3-VL frozen SFT — A-variant arm E

This is arm ngc_anchor_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: ngc_anchor
• Dataset SHA-256: 8890e2fe596e4d78714ce4309f2a59ed9632dd5d656f758c8dc4608d46a9c6bc
• Exposures: 3,588; unique semantic examples: 3,168
• Direct-grounding exposures: 400
• Completion-retention replay exposures: 420
• Image provenance: synthetic AppGen HTML-to-PNG
• 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.