Hcompany/h_cua_perf

H CUA Perf: computer-use agent traces for serving benchmarks

A dataset of agent traces collected by H Company while evaluating its computer-use agent on CUA-Gym desktop tasks, packaged for replay against an LLM inference server with AIPerf (--public-dataset h_cua_perf, see Replaying).

Every record is one real chat-completion request the agent made, with its recorded think time. Replaying the dataset reproduces the request pattern of a tool-calling, screenshot-driven agent: prompts that grow with every step, a bounded number of images per request, tool definitions on every call, and a prefix that is shared with the previous request only up to the point where the agent's context cleaning replaced a screenshot.

What the dataset is

• 477 trajectories, 18,224 requests. 38.2 requests per trajectory on average (median 23), from 1 to 201, the harness capping trajectories at 200 steps. Trajectories were sampled in five step-count buckets, so short and long episodes are about equally represented. A trajectory's length, in the options below, is its number of requests.
• Mooncake trace format. One OpenAI chat-completion request per agent step, with the inter-request delay (the time the agent waited between the previous response and the next request, tool execution included, in milliseconds) and output_length (the recorded completion tokens). There is no absolute timestamp: the trajectories were recorded independently of each other, so the dataset replays at a chosen concurrency or request rate, each session's turns spaced by their recorded delays. session_id is the task id, unique to its trajectory; a trajectory's records are contiguous and trajectories are stored in task-id order.
• Tool calling. The agent drives the desktop through a shell and file tools rather than GUI actions. Each request carries the agent's 10 tool definitions in its tools field and tool_choice: auto in extra. Assistant messages hold tool_calls; observations come back as tool messages.
• Screenshots. The task screenshot arrives with the first user message; later ones are returned inside tool results. This source build keeps one screenshot per request, the latest, and replaces every earlier one in place by the text placeholder [Image omitted by context cleaning], exactly as the agent's own context cleaning did at collection time. A wider window can be set at replay time or derived, see Screenshot window. Screenshots are PNG data URLs; each image part also carries a uuid (xxh3-64 of the data URL) so a replay can send uuid-only references to a server that already holds the image.
• Prompt growth. The prompt holds the system prompt and the full text history of the trajectory (tool calls and tool results included), so prompt length and KV-cache footprint grow roughly linearly with the step index while the number of images stays bounded. When the agent compacted its context, the following requests restart from a summary of the history, so the prompt of those trajectories shrinks once. The average record is 0.5 MB and the largest 2.6 MB.

Files

File	Content
h_cua.jsonl.zst	The source dataset, zstd-compressed (2.5 GB; 9.3 GB decompressed), one screenshot per request
h_cua.meta.json	Manifest: trajectory and request counts, the number of requests of every trajectory (session_turns), the sha256 of the trace file, which AIPerf checks before replaying, and the curation notes (the 23 trajectories excluded from the 500 collected, and why)
trace_processor.py	Derives replay files from the source: screenshot window, trajectory sample, trajectory lengths. Same options as AIPerf's --dataset-filter, see Replaying

Record format

{
  "session_id": "0077a580-d847-507b-a7e1-8cf2b024a23e",
  "delay": 2100,
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": [{"type": "text", "text": "..."}, {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}, "uuid": "..."}, {"type": "text", "text": "..."}]},
    {"role": "assistant", "content": "...", "tool_calls": [{"id": "...", "type": "function", "function": {"name": "shell", "arguments": "{...}"}}]},
    {"role": "tool", "tool_call_id": "...", "content": [{"type": "text", "text": "...\n\n[Image omitted by context cleaning]\n\n..."}]}
  ],
  "tools": [{"type": "function", "function": {"name": "shell", "description": "...", "parameters": {...}}}],
  "output_length": 187,
  "extra": {"tool_choice": "auto"}
}

The first request of a trajectory has no delay. A few requests (25 of 18,224) are not agent steps but the agent's own history-compaction and OCR calls, made without tools; they carry neither tools nor extra and are kept as recorded.

Replaying

AIPerf replays the dataset from the Hub with --public-dataset h_cua_perf; the tutorial Replay H CUA Perf Computer-Use Agent Traces covers the server requirements, the command and its flags. The screenshot window and the trajectory selection are --dataset-filter options, applied at load time. trace_processor.py has the same options, with the same defaults and behaviour, and writes the result to a file for --input-file <file> --custom-dataset-type mooncake_trace replays or other tools:

AIPerf	trace_processor.py	Effect
--dataset-filter n_screenshots=N	--n-screenshots N	Sliding window of the N latest screenshots per request; unset, the published single screenshot. See Screenshot window.
--dataset-filter min_trace_length=N	--min-trace-length N	Drop shorter trajectories; truncations never go below it (default 1).
--dataset-filter max_trace_length=N	--max-trace-length N	Keep each trajectory's first N requests.
--dataset-filter avg_trace_length=X	--avg-trace-length X	Scale every trajectory's length by the same factor until the mean reaches X.
--num-dataset-entries N	--num-traces N --seed S	Number of trajectories: the first N eligible ones in file order for AIPerf, a seeded random sample for the script.

The steps apply in that order: trajectories shorter than the minimum are dropped, the others capped at the maximum, N of them selected, then all of them fitted to the target mean. Trajectories are stored in task-id order, so the first N are not biased by length.

python trace_processor.py --input-path h_cua.jsonl.zst --output-path h_cua-5shots-40.jsonl \
    --n-screenshots 5 --max-trace-length 40

The script reads the compressed file directly and writes a plain JSONL with a .meta.json manifest next to it; python trace_processor.py --help lists every option.

Screenshot window

The source keeps one screenshot per request. A wider sliding window, restoring the trajectory's N latest screenshots into every request, is set with --dataset-filter n_screenshots=N at replay time or --n-screenshots N on trace_processor.py.

A wider window lowers the prefix-cache hit rate. Each request drops the oldest screenshot of the previous window and puts the placeholder in its place, so the prefix shared with the previous request ends at that observation, N-1 steps back instead of one. The screenshots still inside the window are the same bytes (same uuid), so a server can reuse their image embeddings, but their KV entries sit after the divergence and are recomputed. A wider window also lengthens every request by the restored images, so the server's image limit per prompt must be at least N.

How the traces were collected

The tasks are the desktop training tasks of CUA-Gym, run in CUA-Gym's desktop environment. The agent is H Company's computer-use agent driven over a shell and file tools, with a Qwen3.8-27B policy (qwen-3-8-27b), evaluated by CUA-Gym's own task verifiers. Traces were recorded on 2026-09-08 and 2026-09-09. From that collection, the trajectories were sampled by step count as described above; nothing inside a request was altered except the screenshot windowing.

Acknowledgements

The tasks and environments come from CUA-Gym. Task data from CUA-Gym (xlang-ai) is licensed CC BY 4.0. If you use this dataset, please also cite:

@misc{wang2026cuagymscalingverifiabletraining,
      title={CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents},
      author={Bowen Wang and Dunjie Lu and Junli Wang and Tianyi Bai and Shixuan Liu and Zhipeng Zhang and Haiquan Wang and Hao Hu and Tianbao Xie and Shuai Bai and Dayiheng Liu and Que Shen and Junyang Lin and Tao Yu},
      year={2026},
      eprint={2605.25624},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.25624},
}

License

This dataset is released under CC BY 4.0. The task instructions and screenshots derive from CUA-Gym task data, itself CC BY 4.0; keep the attribution above when redistributing.