glm-vision-sft-mix

# glm-vision-sft-mix Merged SFT mix used to train [fable-glm-vision](https://huggingface.co/0xSero/fable-glm-vision) — a 44.8M vision projector that gives GLM-5.2 image understanding with a frozen backbone, following the recipe by Harry Partridge ([original post](https://x.com/part_harry_/status/2077610277571637435)). **151,439 train / 5,354 val** examples. Each row: `{"image": <filename>, "conversations": [user, assistant], "source": <subset>}`. The user turn contains a literal image-span placeholder followed by the instruction; answers are deliberately SHORT (single actions / single phrases) — the recipe found short on-policy answers are required for the SFT grokking event. Coordinates are normalized integers in [0,999], actions in UI-TARS style: `click(start_box=[x,y])`, `type_text(content=...)`, `scroll(...)`, `swipe(...)`, `press_key(...)`, `finished()`. ## Subsets | source | examples | teaches | built from | |---|---|---|---| | `screenshots` | 32.5k | single-step UI grounding (click coordinates) | [agentsea/wave-ui-25k](https://huggingface.co/datasets/agentsea/wave-ui-25k), [showlab/ShowUI-desktop](https://huggingface.co/datasets/showlab/ShowUI-desktop) | | `multistep` | 50.3k | multistep browser/mobile/OS use (instruction + action history -> next action) | [xlangai/aguvis-stage2](https://huggingface.co/datasets/xlangai/aguvis-stage2) (mind2web, aitw, miniwob, guiact-web-multi), Apache-2.0 | | `art` | 74.0k | art & design understanding (artist/style/genre/date, product color/type/season) | [Artificio/WikiArt_Full](https://huggingface.co/datasets/Artificio/WikiArt_Full), [benitomartin/fashion-product-images-small-384x512](https://huggingface.co/datasets/benitomartin/fashion-product-images-small-384x512) | ## Images Images are NOT re-hosted here (licensing differs per source — WikiArt content is research/non-commercial per wikiart.org terms). `image` is the filename produced by the builder scripts; regenerate images from the source datasets with the builders in the [fable-glm-vision GitHub repo](https://github.com/0xSero/fable-glm-vision) (resize to <=300k pixels, dimensions rounded to 28px multiples). ## Credit Method and grokking recipe: Harry Partridge ([@part_harry_](https://x.com/part_harry_)) — [GLM 5.2 With Vision](https://x.com/part_harry_/status/2077610277571637435).