---
license: other
task_categories:
  - image-to-text
tags:
  - gui-agent
  - vision-language
  - supervised-fine-tuning
configs:
  - config_name: ngc-f-qwen25-absolute
    data_files:
      - split: train
        path: data/ngc-f-qwen25-absolute.jsonl
  - config_name: ngc-f-qwen3-normalized
    data_files:
      - split: train
        path: data/ngc-f-qwen3-normalized.jsonl
---

# AppGen NGC SFT v1

Portable, hash-audited MS-SWIFT training data for the NGC-F ablation (plus
NGC-VT only when that optional arm is present in this revision). Images are
procedurally rendered AppGen screenshots from the same 50 training
environments used to construct the SFT rows. They are not real-device or user
screenshots. No public benchmark, AndroidWorld evaluation, held-out
evaluation, or third-party benchmark images are included.

| Config | Arm | Family | Coordinates | Exposures |
|---|---|---|---|---:|
| `ngc-f-qwen25-absolute` | ngc-f | qwen25 | absolute-pixel | 3588 |
| `ngc-f-qwen3-normalized` | ngc-f | qwen3 | normalized-1000 | 3588 |

Each config contains 2,768 canonical navigation rows, 400 direct-grounding
rows, and 420 completion-boundary replay exposures (3,588 total; 3,168 unique
semantic examples). NGC-F uses action-only grounding targets: the assistant
target is a single `<tool_call>` and contains no grounding `<think>`. Canonical
navigation and completion replay rows retain their original
`<think> + <tool_call>` targets.

`source=appgen_direct_grounding_v2` identifies direct-grounding rows.
Navigation and completion rows retain the canonical navigation source tag;
completion duplicates are explicitly identified by `ngcf_v1_replay_role`
(`penultimate_action` or `success_terminate`).

Qwen2.5 coordinates are absolute screenshot pixels. Qwen3 coordinates are
normalized to the integer 0--1000 frame. The two family configs are paired by
semantic ID and differ only where the family coordinate/tool contract requires
it.

## Training reference

The corresponding NGC-F run starts from the original Qwen Instruct backbone,
uses one epoch, global batch 32, learning rate 5e-7, cosine scheduling with 5%
warmup, bf16, ZeRO-2, seed 42, and full LLM tuning while freezing the vision
encoder and multimodal aligner. The manifest records 113 optimizer updates.
These are provenance/reference settings rather than loader defaults.

## Provenance, privacy, and licensing

The screenshots are generated HTML-to-PNG AppGen training-environment renders
with synthetic task content; no real accounts, private conversations, or
captured user/device data are intentionally included. Packaging is fail-closed:
every image must be a PNG inside its row's generated environment and under the
allowed training root, while benchmark/evaluation path tokens are rejected.
Images are content-addressed and deduplicated across family configs.

This dataset is published with `license: other`; redistribution/use remains
subject to the AppGen project and any applicable asset terms. It does not
redistribute model weights. `source_data/` preserves the exact original
MS-SWIFT JSON bytes for audit purposes; those archival files contain original
cluster paths and should not be used directly on another machine.

Images use portable `images/sha256/...` paths and are deduplicated across all
configs in this revision. `bundle_manifest.json` pins every data/image/source-contract
hash. To recover a local MS-SWIFT JSON array:

```bash
python materialize_ms_swift.py \
  --bundle-root . \
  --config ngc-f-qwen25-absolute \
  --output /data/appgen/qwen25_ngc_f_v1.json
```