Guess & Sketch: Language Model Guided Transpilation

Authors: Celine Lee, Abdulrahman Mahmoud, Michal Kurek, Simone Campanoni, David Brooks, Stephen Chong, Gu-Yeon Wei, Alexander M Rush

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test GUESS & SKETCH on three different test sets of assembly transpilation tasks, varying in difficulty, and show that it successfully transpiles 57.6% more examples than GPT-4 and 39.6% more examples than an engineered transpiler.
Researcher Affiliation Academia Cornell University, Harvard University Northwestern University cl923@cornell.edu
Pseudocode Yes Algorithm 1 GUESS & SKETCH Pseudocode
Open Source Code No The paper states "The resulting dataset is shared on Hugging Face" and "All resulting models are shared on Huggingface", and links to a baseline's code, but it does not provide an explicit statement or link for the source code of the GUESS & SKETCH methodology.
Open Datasets Yes Training data is composed of 307,916 ARMv8 and RISC-V assembly file pairs compiled from C code files from The Stack (Kocetkov et al., 2022). The resulting dataset is shared on Hugging Face3. 3https://huggingface.co/datasets/celinelee/paired arm risc
Dataset Splits No The paper describes the training dataset and separate test datasets, but it does not provide specific train/validation/test splits for the main training data or how the models were validated during training.
Hardware Specification Yes All language models are trained on one NVIDIA RTX A6000 GPU.
Software Dependencies No The paper mentions using "Rosette (Torlak & Bodik, 2013)" and "Z3 (de Moura & Bjørner, 2008) SMT solver" but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We use confidence threshold γ = 0.9 and Table 4: Training details for language models used. which includes L.R., Batch, No. Steps, Lo RA r, Lo RA Modules, Quant. values.