Synthesize, Partition, then Adapt: Eliciting Diverse Samples from Foundation Models

Authors: Yeming Wen, Swarat Chaudhuri

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

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the effectiveness of our approach, we conduct experiments on a range of tasks in both the code generation and natural language understanding domains. We evaluate our method on the Human Eval (Chen et al., 2021) and MBPP (Austin et al., 2021) datasets for code generation, as well as several natural language understanding tasks.
Researcher Affiliation Academia Yeming Wen & Swarat Chaudhuri Department of Computer Science The University of Texas at Austin
Pseudocode No The paper includes Figure 3 as an illustration of the SPA framework, but it is a diagram, not structured pseudocode or an algorithm block.
Open Source Code Yes The code is included in the supplementary material.
Open Datasets Yes For the synthetic dataset, we utilize the OSS-Instruct dataset (Wei et al., 2023)... For the synthetic dataset, we use Platypus (Lee et al., 2023)...
Dataset Splits No The paper describes partitioning the synthetic datasets into K groups for training different model adaptations, but it does not specify explicit training/validation/test splits for the overall experimental evaluation of the models.
Hardware Specification Yes For the code generation experiments, we use a machine with 3 A100 40GB GPUs and train each partition for 400 steps... On the OSS-Instruct dataset, this takes roughly 5 hours using a single A100 80GB GPU.
Software Dependencies No The paper mentions software components like 'LORA', 'BM25', 'Python' and 'torch-influence' but does not specify their version numbers.
Experiment Setup Yes For the code generation experiments, ... The base model is Code LLa MA-7B-Python, and we use bf16 precision to accelerate training. The per-device train batch size is set to 1, with a gradient accumulation step of 20. We use a learning rate of 2e-4 with a cosine learning rate scheduler and 20 warmup steps. For the LORA hyperparameters, we use a rank (r) of 16, an alpha of 16.