PlanFM Resources
Repositories and datasets connected to PlanFM research on compact foundation models for planning-like tasks, state-centric representations, learned transition models, and downstream planning evaluation.
Model Compression - Quantization
Code for quantization experiments associated with compact state-centric planning models and efficient deployment of planning foundation model components.
Tokenizer Design
Code for representing symbolic planning states as learning-ready tokens, including tokenization choices used to study generalization in learned transition models.
Downstream Applications - Plan Validity
Code for the downstream plan-validity task, where state-centric learning is used to assess whether generated plan traces satisfy planning constraints.
Model Architecture and Training
Official implementation for learning transition dynamics over state representations and decoding executable plans through symbolic successor selection.
Data Preparation
FABLE evaluates language models on data-flow analysis over procedural text, including planning scenarios, travel routes, and recipes. The FABLE+ dataset is available on Hugging Face, with code and benchmark generation support in the GitHub repository.