# Cognyzer OSWorld sample pack for Z.ai

Ten computer-use tasks from Cognyzer. Six Ubuntu tasks include scored model rollouts. Four Windows and macOS tasks ship as SFT only: a passing human trajectory, no model runs.

Every task is an executable environment. A run passes only if the artefact left inside the virtual machine satisfies the evaluator.

Contact: Pragnasya S, Co-founder and CTO, Cognyzer, Hyderabad. team@cognyzer.com

## Layout

Ubuntu stays under `ubuntu/`, Windows under `windows/`, macOS under `macos/`.

```
osworld-sample-zai/
├── README.md
├── ubuntu/
├── windows/
└── macos/
```

## Per-task structure

Ubuntu tasks (RL, with model rollouts):

```
thunderbird-c37b60/
├── task.json                     definition, environment config and evaluator
├── sft/                          human reference trajectory
│   ├── trajectory.jsonl          one record per executed action
│   ├── result.txt                1.000000
│   └── step_0_before.png ...
└── runs/                         scored rollouts, grouped by model
    ├── <model>/run_01/           result.txt, trajectory.jsonl, screenshots
    ├── <model>/run_02/
    └── <model>/run_03/
```

Windows and macOS tasks (SFT only, no `runs/`, no `model_pass_rate`):

```
gsheets/
├── task.json                     definition, environment config and evaluator
├── assets/                       seed files used at setup
└── sft/                          human reference trajectory
    ├── trajectory.jsonl          one record per executed action
    ├── result.txt                1.000000
    └── step_0.png ...
```

Every `sft/result.txt` is `1.000000`. That is the evaluator score for the human trajectory. Model `result.txt` files under `runs/` are `1.000000` or `0.000000` depending on the rollout.

`task.json` on Windows and macOS does not include `model_pass_rate`, `annotator_hints`, or `knowledge_points`. Ubuntu keeps `model_pass_rate` because those six tasks were scored.

## Ubuntu RL (scored)

Harness is identical across rollouts so the model is the only variable.

| Setting | Value |
| --- | --- |
| Environment | Ubuntu 22.04, 1920x1080 |
| Observation | screenshot only |
| Action space | pyautogui, generated by the model |
| Memory | the last 3 screenshots |
| Step ceiling | 50 for Claude, 60 for GLM |
| Settle time | 2 seconds after each action |
| Runs per model | 3 |

Pass rate is the share of successful runs out of three, so `1 1 0` is 0.667.

Models that report coordinates on a 0 to 1000 grid (GLM V-series) have their output rescaled to real pixels before execution. Without that step every click lands in the top left eighth of the screen.

| Task | Application | Model | Runs | Pass rate |
| --- | --- | --- | --- | --- |
| `thunderbird-c37b60` | Thunderbird | claude-sonnet-4-5-20250929 | `1 1 0` | 0.667 |
|  |  | glm-5v-turbo | `0 0 0` | 0.0 |
| `vlc-5e94af` | VLC | claude-sonnet-4-5-20250929 | `1 1 0` | 0.667 |
|  |  | glm-5v-turbo | `0 0 0` | 0.0 |
| `vs-code-4b6f82` | VS Code | claude-sonnet-4-5-20250929 | `1 0 0` | 0.333 |
|  |  | glm-5v-turbo | `0 0 0` | 0.0 |
| `gimp-ef7602` | GIMP | glm-5v-turbo | `0 0 0` | 0.0 |
| `libreoffice-writer-7cfdbc` | Writer | glm-5.3-flash | `0 0 0` | 0.0 |
| `libreoffice-calc-e4af30` | Calc | glm-5.3-flash | `0 0 0` | 0.0 |

## Windows and macOS (SFT)

| Folder | OS | Application | What the human does |
| --- | --- | --- | --- |
| `windows/gsheets` | Windows | Google Sheets | Reykjavik weather workbook |
| `windows/gdocs` | Windows | Google Docs | Thesis workshop document edits |
| `macos/onlyoffice` | macOS | ONLYOFFICE | Fleet insurance audit columns and validation |
| `macos/chrome` | macOS | Chrome / Sheets | Q2 ridership derived columns |

## Reading a run

Each line of `trajectory.jsonl` is one executed action. Several lines can share a step number when the model issued a batch of actions in one turn.

No model reasoning is included. The response field holds the emitted action only.

Ubuntu seed downloads in `task.json` point at `huggingface.co/datasets/Cognyzer/osworld-ubuntu`. Windows and macOS seed files are in each task's `assets/` folder.

This pack is confidential Cognyzer material, shared with Z.ai for evaluation of a proposed computer-use training and evaluation collaboration.