HPAI-BSC/NotSoTiny-25-12-CVDP

This is a version of NotSoTiny-25-12 benchmark modified to work with Nvidia's CVDP framework

If you plan to run this benchmark via CVDP framework, proceed with this version. Otherwise refer to the main dataset for additional info: HPAI-BSC/NotSoTiny-25-12

Subsets and Shuttles

The default dataset contains all NotSoTiny-25-12 Tiny Tapeout shuttles combined onto a single dataset comprising 1114 total tasks. However, you can also access individual shuttles as subsets.

Subset	# Tasks	Launched date	Tiny Tapeout source
default	1,114	-	-
tt06	108	2024-01-30	https://tinytapeout.com/chips/tt06/
tt07	177	2024-04-22	https://tinytapeout.com/chips/tt07/
tt08	196	2024-06-10	https://tinytapeout.com/chips/tt08/
tt09	250	2024-09-07	https://tinytapeout.com/chips/tt09/
tt10	214	2025-03-12	https://tinytapeout.com/chips/ttihp25a/ (and ttihp0p2)
ttsky	169	2025-06-27	https://tinytapeout.com/chips/ttsky25a/

Run it with CVDP

1. Clone the CVDP framework:

$ git clone https://github.com/NVlabs/cvdp_benchmark.git
$ cd cvdp_benchmark/
$ uv init && uv add -r requirements.txt

2. Download the dataset shuttles:

$ git lfs install
$ git clone https://huggingface.co/datasets/HPAI-BSC/NotSoTiny-25-12-CVDP
$ ls NotSoTiny-25-12-CVDP/shuttles/
    tt06.jsonl   tt07.jsonl   tt08.jsonl   tt09.jsonl   tt10.jsonl   ttsky.jsonl
$ cp .env.example .env && echo "OSS_SIM_IMAGE=ggcr0/turtle-eval:2.3.4" >> .env

3. [Optional] Validate with golden solutions:

$ uv run run_benchmark.py -f NotSoTiny-25-12-CVDP/shuttles/tt06.jsonl

4. Run the benchmark (inference + eval):

$ export OPENROUTER_API_KEY=sk-or-v1-... 
$ uv run run_samples.py \
       -f NotSoTiny-25-12-CVDP/shuttles/tt06.jsonl \
       -l \
       -m mistralai/codestral-2508 \
       -c examples/openrouter_factory.py \
       -n 5 \
       -k 1

You can repeat this process for any other shuttle present in the shuttles/ dir.

We also recommend the usage of the -t <workers> flag to speed-up the process.

Additional Usage

from datasets import load_dataset

# Load the complete dataset with all shuttles (1,114 tasks)
ds = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", split="test")

# Or load a specific shuttle
ds_tt06  = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", "tt06",  split="test") # 108 tasks
ds_tt07  = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", "tt07",  split="test") # 177 tasks
ds_tt08  = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", "tt08",  split="test") # 196 tasks
ds_tt09  = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", "tt09",  split="test") # 250 tasks
ds_tt10  = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", "tt10",  split="test") # 214 tasks
ds_ttsky = load_dataset("HPAI-BSC/NotSoTiny-25-12-CVDP", "ttsky", split="test") # 169 tasks

License

The dataset is released under the Apache License 2.0.

Citation Information

@misc{ghorab2025notsotinylargelivingbenchmark,
      title={NotSoTiny: A Large, Living Benchmark for RTL Code Generation}, 
      author={Razine Moundir Ghorab and Emanuele Parisi and Cristian Gutierrez-Gomez and Miquel Albert\'i-Binimelis and Miquel Moreto and Dario Garcia-Gasulla and Gokcen Kestor},
      year={2025},
      eprint={2512.20823},
      archivePrefix={arXiv},
      primaryClass={cs.AR},
      url={[https://arxiv.org/abs/2512.20823](https://arxiv.org/abs/2512.20823)}, 
}

Acknowledgements

The HPAI team behind NotSoTiny would like to thank the Tiny Tapeout community for the open source efforts, which made possible this contribution. Special thanks to Matt Venn for his support.

We would also like to thank the CVDP team at NVIDIA for developing and open-sourcing the CVDP framework and to the Si2 Coalition Extend and Expand working group.