CheckmateRTL-9B / README.md
ggcristian's picture
Update model card
2162a09 verified
|
Raw History Blame Contribute Delete
4.63 kB
---
base_model: Qwen/Qwen3.5-9B
library_name: transformers
model_name: CheckmateRTL-9B
tags:
- verilog
- systemverilog
- design-verification
- testbench-generation
- trl
- sft
- rlvr
- grpo
datasets:
- AI-Institute-BSC/SystemVerilog-Distill-86k
- AI-Institute-BSC/SystemVerilog-RL-12k
pipeline_tag: text-generation
---
# CheckmateRTL-9B
A fine-tune of [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) for SystemVerilog design verification: RTL → Test-Plan, (RTL, Test-Plan) → Testbench, and adding functional checks to an existing stimulus testbench. Training combines Gemma-4-31B distillation with reinforcement learning from verifiable rewards, targeting testbenches that pass on the golden DUT and reject faulty variants.
🤗 **Models**
- [CheckmateRTL-27B](https://huggingface.co/AI-Institute-BSC/CheckmateRTL-27B)
- [CheckmateRTL-9B](https://huggingface.co/AI-Institute-BSC/CheckmateRTL-9B) ← you are here
🤗 **Datasets**
- [SystemVerilog-Distill-86k](https://huggingface.co/datasets/AI-Institute-BSC/SystemVerilog-Distill-86k)
- [SystemVerilog-RL-12k](https://huggingface.co/datasets/AI-Institute-BSC/SystemVerilog-RL-12k)
## Evaluation on NVIDIA CVDP benchmark
The results below are reported in the paper. Pass@1 is estimated from five samples per problem. All models are evaluated with vLLM under the same scoring flow.
<p align="center">
<img src="assets/cvdp_stimulus_checking.png" width="100%" alt="CVDP non-agentic pass@1. Testbench stimulus (cid012): CheckmateRTL-27B 37.0%, GLM-5.2 34.0%, CheckmateRTL-9B 29.5%, Gemma-4-31B 24.8%, LLM4Cov 22.4%, DeepSeek-V4 22.1%. Testbench checking (cid013) versus parameter count: CheckmateRTL-27B 36.2% and CheckmateRTL-9B 23.0%, above every baseline including Kimi-K2.5 (1T) at 23.4%." />
</p>
### CVDP v1.1.0 Non-agentic design verification
| Model | Stimulus (cid012) | Checking (cid013) | Assertion (cid014) | Debug (cid016) |
| --- | ---: | ---: | ---: | ---: |
| Kimi-K2.5 | 22.1% | 23.4% | 36.1% | 64.6% |
| GLM-5.2-FP8 | 34.0% | 14.0% | 33.1% | 44.6% |
| DeepSeek-V4-Flash-0731 | 22.1% | 18.5% | 28.1% | **65.1%** |
| Gemma-4-31B | 24.8% | 17.7% | 33.1% | 63.4% |
| Qwen3.6-27B | 19.4% | 15.1% | 22.4% | 58.9% |
| Qwen3.5-9B | 3.0% | 1.9% | 4.8% | 22.9% |
| **CheckmateRTL-9B** | 29.5% | 23.0% | 15.8% | 29.1% |
| **CheckmateRTL-27B** | **37.0%** | **36.2%** | **42.4%** | 60.6% |
### CVDP v1.1.0 Heavy Agentic design verification
Although training uses single-turn tasks, the models are also evaluated with OpenCode as a coding-agent harness, without a custom evaluator-feedback loop.
| Model | Stimulus (cid012) | Checking (cid013) | Assertion (cid014) | Debug (cid016) |
|---|---:|---:|---:|---:|
| Kimi-K2.5 | 37.5% | 40.0% | 37.1% | 54.6% |
| GLM-5.2-FP8 | 52.5% | 22.9% | **60.0%** | **58.6%** |
| DeepSeek-V4-Flash-0731 | **65.0%** | 34.3% | 51.4% | 57.9% |
| Gemma-4-31B | 17.5% | 14.3% | 22.9% | 31.8% |
| Qwen3.6-27B | 32.5% | 17.1% | 20.0% | 46.8% |
| Qwen3.5-9B | 5.0% | 0.0% | 5.7% | 18.6% |
| **CheckmateRTL-9B** | 25.0% | 11.4% | 20.0% | 24.6% |
| **CheckmateRTL-27B** | **65.0%** | **42.9%** | 31.4% | 46.4% |
### AutoEval-ECov
AutoEval-ECov reformulates 156 VerilogEval problems as a testbench stimulus-generation benchmark. Given a specification and the DUT source, the model must generate a testbench that reaches 100% target coverage.
| Model | Pass@1 |
|---|---:|
| Kimi-K2.5 | 37.1% |
| GLM-5.2-FP8 | 81.5% |
| DeepSeek-V4-Flash-0731 | 43.5% |
| Gemma-4-31B | 77.7% |
| Qwen3.6-27B | 56.7% |
| Qwen3.5-9B | 3.2% |
| **CheckmateRTL-9B** | 79.5% |
| **CheckmateRTL-27B** | **84.7%** |
### Evaluation scope
[CVDP](https://github.com/NVlabs/cvdp_benchmark) v1.1.0 stimulus, checking, and assertion scoring is ported from commercial EDA tools to Verilator. AutoEval-ECov is evaluated with the same open-source approach; debugging retains CVDP's original open-source flow. These results are not directly comparable with scores from the commercial flow. Assertion scores are upper-bound estimates because Verilator does not fully support assertion coverage.
## Citation
```bibtex
@inproceedings{gutierrezgomez2027coverage,
author = {Cristian Gutierrez-Gomez and Arnau Ayguade and Bernat Homs and Emanuele Parisi and Razine Moundir Ghorab and Gokcen Kestor and Miquel Moreto and Dario Garcia-Gasulla},
booktitle = {2027 Design, Automation \& Test in Europe Conference (DATE)},
title = {From Coverage to Checking: A Post-Training Recipe for SystemVerilog Testbench Stimulus and Checking},
year = {2027},
url = {}
}
```
## Contact
[cristian.gutierrez@bsc.es](mailto:cristian.gutierrez@bsc.es)