p-doom/cuagym-qwen35-rollouts

---

CUA-Gym Qwen3.5 rollouts

55,139 computer-use trajectories produced by rolling Qwen3.5 over the full CUA-Gym task set inside real Ubuntu desktop VMs, with every episode scored by the task's own reward function.

Two models, both with reasoning enabled: Qwen3.5-35B-A3B over the whole corpus four times, and Qwen3.5-9B over most of it twice.

Dataset Summary

Computer-use agents are trained on demonstrations, and demonstrations are expensive: OSWorld-style tasks need a booted desktop, a scripted initial state, and a grader that inspects the machine afterwards. CUA-Gym supplies 10,910 such tasks. What it does not supply is trajectories.

This dataset is the trajectories. Each task was set up in a KVM virtual machine, handed to a Qwen3.5 model as a screenshot, and driven step by step — screenshot in, computer_use tool call out, pyautogui executed in the guest — until the model called terminate or hit an 80-step budget. The task's own reward.py then ran inside the guest and scored the final machine state.

Every rollout is included, whatever it scored. This is a distillation corpus, not a curated success set: a run that flails for eighty steps and fails is still a record of how a 35B model attacks a spreadsheet. Filter on reward yourself if you want the successes.

Key Features

• 55,139 rollouts over 9,534 distinct tasks, 3.2M screenshots at 1920×1080
• Real desktops, not simulators — LibreOffice Calc/Writer/Impress, VS Code, PDF viewers, VLC, GIMP, plus 31 self-hosted mock web apps, all in KVM VMs
• Graded by each task's own reward.py, executed in-guest; dense partial credit in [0, 1], not just pass/fail
• Two model scales on identical tasks, so 35B and 9B are directly comparable
• Cursor state recorded per step, which is what makes conversion to relative (move_rel) action spaces reproducible

Dataset Statistics

Split	Model	Rollouts	k	Screenshots	Reward > 0	Reward = 1.0	Mean reward
Total		55,139		3,245,904	19.4%	4.1%	0.107
p1_35b_think	Qwen3.5-35B-A3B	38,136	4.00	2,346,434	22.8%	5.0%	0.125
p2_9b_think	Qwen3.5-9B	17,003	1.78	899,470	15.1%	2.2%	0.070

k is samples per task, computed from the data rather than from intent. The 35B split covers all 9,534 tasks exactly four times. The 9B split is a partial pass: its cluster reservation ended mid-run, so it covers the same task list at k ≈ 1.78 rather than a round number. Nothing was cherry-picked when it stopped — tasks are ordered repeat-major, so the truncation is uniform across the corpus, not concentrated in one app family.

The 35B model roughly doubles the 9B model's success rate on identical tasks (22.8% vs 15.1% scoring above zero, 5.0% vs 2.2% scoring a perfect 1.0).

Task composition

Family	Tasks	Examples
Desktop office	6,071	LibreOffice Calc, Writer, Impress
Desktop other	1,958	VS Code, PDF, VLC, GIMP
Web / cross-app	1,505	Slack, Gmail, Notion, Salesforce, GitHub, Jira mocks

Web tasks reference mock applications by URL. They were run against a local CUA-Gym-Hub deployment (31 apps served per compute node) with the endpoint placeholders materialized to that instance.

Format

WebDataset-style shards plus one metadata file per split:

p1_35b_think/trajectories.jsonl      one JSON record per rollout
p1_35b_think/screenshots-*.tar       <task_id>/step_NNN.png
p2_9b_think/...

Each record carries the instruction, the final reward, and a list of steps. Per step: the raw model output, the parsed action, the executed pixel coordinate, the cursor position beforehand, and a reference to the screenshot the model saw (shard + member).

{
  "task_id": "0018392e-…__r2",
  "instruction": "Set the row height of the header row to 0.8 cm.",
  "reward": 1.0,
  "screen": [1920, 1080],
  "steps": [
    {
      "step": 0,
      "screenshot": "screenshots/step_000.png",
      "shard": "screenshots-0041.tar",
      "member": "0018392e-…__r2/step_000.png",
      "cursor_before": [1728, 972],
      "action": "left_click",
      "coordinate_screen": [412, 233],
      "raw_action_args": {"action": "left_click", "coordinate": [215, 216]},
      "assistant_raw": "…</think>\n<tool_call>{…}</tool_call>"
    }
  ]
}

Three things worth knowing before you train on it:

• Coordinates appear in two spaces. The model emits Qwen-native values normalized to 0–1000 (raw_action_args.coordinate); coordinate_screen is what was actually clicked in pixels. Treating the former as pixels puts every click in the top-left corner and silently produces a corpus of failures.
• cursor_before is load-bearing, not diagnostic. Converting these absolute actions into relative move_rel deltas requires knowing where the pointer was, and it cannot be reconstructed afterwards — dialogs and focus changes move the cursor without the agent asking.
• reward: null means the grader produced no parseable score (~5% of rollouts), usually a missing in-guest dependency. Those rollouts are kept and flagged rather than imputed as zero.

Generation setup

Environment	OSWorld DesktopEnv, native QEMU/KVM, Ubuntu guest, 1920×1080
Serving	SGLang, bf16, TP=4 per node, 16×4 NVIDIA A100-40GB
Sampling	temperature 0.7, max_tokens 3072, reasoning enabled
Step budget	80 actions per episode
Concurrency	256 simultaneous VMs (16 nodes × 16 workers)

The 80-step budget is deliberate. At the 40 steps used in earlier pilots, 39% of episodes were still truncated mid-task; even at 80 the median episode uses the full budget, so the underlying distribution of task lengths extends further still. Truncated episodes are labelled (terminated: false) and kept.

Attribution

Tasks come from CUA-Gym by XLANG Lab; mock web applications from CUA-Gym-Hub. The environment is a fork of OSWorld. Models are Qwen3.5 by Alibaba. Compute was provided by the HoreKa supercomputer at KIT.

If you use this dataset, please cite the underlying task set alongside it.