joooelw/ToM-PBM-3B
joooelw/ToM-PBM-3B is a 3.2 billion parameter instruction-tuned causal language model, fine-tuned from Meta Llama-3.2-3B-Instruct. Developed as part of the DEL-ToM project, this model introduces a Dynamic Epistemic Logic-based framework for modeling and evaluating theory-of-mind reasoning in large language models. It is specifically optimized for advanced cognitive tasks involving understanding and predicting the mental states of others.
Loading preview...
Model Overview
joooelw/ToM-PBM-3B is a 3.2 billion parameter language model, fine-tuned from meta-llama/Llama-3.2-3B-Instruct. This model is a core component of the DEL-ToM project, which focuses on enhancing Theory-of-Mind (ToM) reasoning in large language models through a novel Dynamic Epistemic Logic-based framework. The project's paper, "DEL-ToM: Inference-Time Scaling for Theory-of-Mind Reasoning via Dynamic Epistemic Logic," provides further details on its theoretical underpinnings.
Key Capabilities
- Advanced Theory-of-Mind Reasoning: Specifically designed and fine-tuned to model and evaluate complex cognitive processes related to understanding beliefs, intentions, and knowledge of others.
- Dynamic Epistemic Logic Framework: Utilizes a sophisticated logical framework to represent and reason about epistemic states, offering a structured approach to ToM tasks.
- Llama-3.2 Base: Built upon the robust Llama-3.2-3B-Instruct architecture, providing a strong foundation for language understanding and generation.
Training Details
The model was trained using Axolotl, with a learning rate of 2e-06, a batch size of 1 (accumulated to 4), and a cosine learning rate scheduler over 1 epoch. It leverages paged_adamw_32bit optimizer and flash_attention for efficient training. The training dataset used was /home/ubuntu/LLM-inference/yuheng-project/tts/ToM_PRM_final.jsonl, configured for Llama3 conversation format with a sequence length of 8192 tokens.