CodeV-R1 Dataset

This dataset contains the data required for training CodeV-R1, a specialized model for Verilog code generation.

Dataset Composition

📁 Files

• codev_r1_sft.jsonl: Supervised fine-tuning dataset for distilling CodeV-R1-Distill from Qwen2.5-Coder-7B-Instruct
• codev_r1_rl_train.parquet: Reinforcement learning training dataset for obtaining CodeV-R1 from CodeV-R1-Distill
• codev_r1_rl_val.parquet: Validation dataset for RL training

🛠️ Data Construction Details

Supervised Fine-Tuning Data (codev_r1_sft.jsonl)

Our distillation pipeline begins with Verilog code samples collected from GitHub:

1. Instruction Generation: DeepSeek-V3 summarizes these code snippets to produce corresponding instructions xᵢ
2. Response Generation: DeepSeek-R1 generates complete responses containing "thought" and Verilog code
3. Quality Filtering: We apply rigorous curation:
   • Retain only instructions where baseline models (Qwen2.5-Coder-7B/32B-Instruct) fail functional verification
   • Ensure synthesizability of collected Verilog code with Yosys
   • Remove samples with Rouge-L similarity > 0.5 to existing benchmarks (VerilogEval v1/v2, RTLLM v1.1/v2) to prevent contamination

This process yields 87K high-quality NL-thought-code triples for supervised distillation.

Reinforcement Learning Data (codev_r1_rl_train.parquet, codev_r1_rl_val.parquet)

Our RL data curation focuses on three key criteria: solvability, challenge, and error-free quality:

1. Equivalence Validation: We identify problems where DeepSeek-R1 generates code matching the original GitHub code, while both Qwen2.5-Coder-7B and 32B models fail
2. Difficulty Enhancement: CodeV-R1-7B-Distill generates five code variants per question, excluding cases where all variants match the golden code (collected from Github)
3. Quality Assurance: We prioritize problems that require reasoning and ensure validity between instructions and the golden code

Through this rigorous selection, we curate 3.1K high-quality examples for reinforcement learning, specifically designed to enhance the model's reasoning capabilities on challenging Verilog problems.

📊 Key Statistics

• Initial triples generated: ~150K
• After quality filtering: 87K (SFT)
• RL training examples: 3.1K (3K in codev_r1_rl_train.parquet and 100 in codev_r1_rl_val.parquet)

🎯 Purpose

This dataset enables:

• Supervised distillation from Qwen2.5-Coder-7B-Instruct to CodeV-R1-Distill
• Reinforcement learning fine-tuning to produce the final CodeV-R1 model
• Enhanced reasoning capabilities for Verilog code generation tasks