Exploring the LLM Journey from Cognition to Expression with Linear Representations

Authors: Yuzi Yan, Jialian Li, Yipin Zhang, Dong Yan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our study includes a comprehensive series of experiments and analyses carried out during the Pretraining, SFT, and RLHF phases of the Baichuan-7B and Baichuan33B. We carry out our quantification experiments using four standard benchmark datasets: Openbook QA (Mihaylov et al., 2018), Common Sense QA (Talmor et al., 2018), RACE (Lai et al., 2017) and ARC (Clark et al., 2018).
Researcher Affiliation Collaboration 1Baichuan AI 2Tsinghua University.
Pseudocode Yes Algorithm 1 Cognitive capability quantification
Open Source Code No The paper states: 'Notably, Baichuan7B is an open-source model, whereas Baichuan-33B is a closed-source model.' This refers to the models used in the study, not the authors' own code for their methodology. No statement is made about releasing the code for the research presented in the paper.
Open Datasets Yes We carry out our quantification experiments using four standard benchmark datasets: Openbook QA (Mihaylov et al., 2018), Common Sense QA (Talmor et al., 2018), RACE (Lai et al., 2017) and ARC (Clark et al., 2018). Table 3. Size for each datasets. DATASET NAME TRAINSET SIZE TESTSET SIZE
Dataset Splits Yes Table 3. Size for each datasets. DATASET NAME TRAINSET SIZE TESTSET SIZE
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. It mentions 'computational power worth billions is used daily' for training LLMs generally, but not for their specific experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies used in the experiments (e.g., programming languages, libraries, or frameworks).
Experiment Setup Yes The SFT phase involved training over 4 epochs, each with 1M tokens. For RLHF, we implement the Proximal Policy Optimization (PPO) strategy, as elaborated in Achiam et al. (2023). Table 4. Direct token generation hyperparameters. TERM PARAMETER. TEMPERATURE 1.2, TOP P 0.9, TOP K 50, MAX TOKENS 2048, REPETITION PENALTY 1.05.