The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 84, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Micius-Flash Bench
Benchmark assets for the Micius-Flash technical report that are too large for git. The evaluation code, the one-click scripts and the smaller benchmarks (SemiBench, PCB Agent inputs) live in the GitHub repository Zhiman-BJ/Micius-Flash.
| Benchmark | Where |
|---|---|
| SemiBench (167 questions) | GitHub data/semibench/ |
| PCB Agent (30 tasks) | GitHub data/pcb_agent/ + container images ghcr.io/zhiman-bj/pcb-agent-env:v34, ghcr.io/zhiman-bj/pcb-agent-verifier:kicad-10.0.6 |
| Semiconductor Agent (197 tasks) | this dataset |
Semiconductor Agent
Semiconductor Agent measures whether a model can finish realistic engineering work by calling domain tools in an executable environment: reconstructing provenance in an AiiDA archive, checking timing sign-off reports, reducing EELS / 4D-STEM / CβV measurement data, auditing fab equipment logs, running RTL simulations, and so on.
| Tasks | 197 held-out tasks, one per environment |
| Environment | record database (SQLite) + workspace files + task-specific tools (median 26, up to 88), each tool running in a pinned Python software profile |
| Grading | frozen per-task rubric applied by an independent Codex verifier |
| Metric | pass rate over all 197 tasks (fixed denominator) |
Files
semiconductor_agent/
βββ tasks.jsonl # 197 rows: case_id, environment_id, task_id, package_id, software_profile
βββ manifest.json
βββ archives/
βββ environments.tar.zst.part-000 # environments/ (task, records, workspace) + rubrics/ (frozen verifier rubrics)
βββ runtimes.tar.zst.part-000, -001 # source_tree/: tool delivery packages, software profiles (uv venvs), interpreters
βββ SHA256SUMS
About 39 GB compressed and about 145 GB after extraction (identical files are hard-linked inside the archives). Linux x86-64 only.
Usage
Do not unpack the archives by hand. The setup script verifies, extracts and relocates them (the software profiles contain absolute paths that are rewritten to your location):
git clone https://github.com/Zhiman-BJ/Micius-Flash && cd Micius-Flash
pip install -e .
export MICIUS_DATA=/path/with/enough/space
bash scripts/download_data.sh semiconductor_agent
Then follow docs/semiconductor_agent.md to run the agent and the verifier.
License
The task data, environments and rubrics are released under Apache 2.0. The tool runtimes bundle third-party open-source packages and interpreters, which remain under their own licenses.
Citation
@misc{micius2026flash,
title = {Micius-Flash: Lightweight Agentic Intelligence for Engineering},
author = {Micius Team},
year = {2026},
eprint = {2611.10015},
archivePrefix = {arXiv},
primaryClass = {cs.SE},
url = {https://arxiv.org/abs/2611.10015}
}
- Downloads last month
- 40