PlanLLM: Task-Oriented Conversational Assistant
PlanLLM is a 7 billion parameter conversational assistant developed by NOVA-vision-language, specifically designed to guide users through complex, multi-step tasks. Built upon the Vicuna-7B-v1.1 architecture, this model excels at providing step-by-step assistance and answering relevant questions for activities such as following recipes or completing DIY projects.
Key Capabilities
- Task Guidance: Assists users from beginning to end in completing structured tasks like recipes.
- Contextual Understanding: Capable of answering related and relevant user requests throughout a task.
- Conversational Flow: Trained on synthetic user-system dialogues to maintain coherent and helpful interactions.
Training Methodology
PlanLLM was fine-tuned using a two-stage process:
- Supervised Fine-Tuning (SFT): Initial training on 10,000 synthetic dialogues, covering 1,000 unique recipes, using FSDP on 4 A100 GPUs.
- Direct Preference Optimization (DPO): Further refinement on 3,000 dialogues (from the same 1,000 recipes) using LoRA on a single A100 GPU to align with preferred conversational patterns.
Dataset
The model's training data consists of synthetic user-system dialogues. User utterances were derived from real Alexa users interacting with the TWIZ taskbot, mapped to specific intents. System responses were generated using templates, external knowledge, and other Large Language Models, creating realistic dialogue flows.
Research Paper
This model is associated with the paper "Plan-Grounded Large Language Models for Dual Goal Conversational Settings" accepted at EACL 2024. Read the paper here.
License
PlanLLM operates under the same non-commercial Apache 2.0 license as its base model, Vicuna.