Training contract
- Base:
Qwen/Qwen3-VL-8B-Instruct at commit 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
- Dataset:
luca0621/appgen-sft-ngc-v1, config ngc-f-qwen3-normalized, revision 769ea99dbc4ff190048ae0db37eb6310dba595e0
- Dataset variant:
policy_clean_grounding2x
- Dataset SHA-256:
ec34c34e50ee78ce426baaa58dcd71427cf4309688f2b6a79285fcfa319fd15d
- Exposures: 3,982; unique semantic examples: 3,165
- Direct-grounding exposures: 800
- Completion-retention replay exposures: 417
- Coordinates: normalized 0–1000
- Prompt variant:
policy_v1; SHA-256 f47248dbafaff3e04750b75952c2ccab2dfe325705e9b850ec45452099e4f5e9
- Optimizer: full-language AdamW, learning rate
5e-7, cosine schedule, 5% warmup
- Batch: microbatch 4 × two GPUs × gradient accumulation 4 = global batch 32
- Epochs: 1; optimizer updates: 125
- Frozen: visual encoder and aligner; trainable: language model and LM head
- Swift's
hermes loss-scale preset was used (the sweep's tool2x arm).
Only the final checkpoint (checkpoint-125) is published in
this repository. The exact training system prompt is included as
appgen_system_prompt.txt; run_manifest.json records hashes and verification
evidence. Intermediate 25-step checkpoints remain local for controlled eval.
Policy-cleaning disclosure (arm D)
This arm starts from the 3,988-row grounding-2× variant. It removes six
publisher-identified mislabeled terminate(success) exposures: three original
rows and their three completion-replay rows. Selection uses the publisher
task_id plus publisher idx, not mutable row positions. It also removes the
undeclared time: 1 argument from exactly two long_press tool calls; only
those tool-call JSON objects were reserialized. The resulting dataset has
3,982 rows, 3,165 unique semantic examples, and no remaining undeclared
time fields.
Data provenance and limitations
The images are synthetic AppGen HTML-to-PNG screenshots from 50 generated
training environments. They are not real-user or real-device captures. The
pinned dataset declares no AndroidWorld, held-out, or public grounding
benchmark images. No AndroidWorld or real-world benchmark score is claimed in
this card; evaluate all four arms under the same decoding and benchmark setup.
This model is for Android visual-agent research, not safety-critical autonomous
deployment.