How para-lost/autopresent works
Project Type and Purpose
• This is an open‐source research codebase implementing AutoPresent, a system for automatic slide generation.
• It provides:
• A core library (SlidesLib) for programmatically creating slides
• A pipeline to generate Python “programs” (via LLM inference), refine them, execute them to build slides
• A training component (autopresent) for fine-tuning Llama models on slide‐creation tasks
• A benchmarking suite (slidesbench) with datasets and reference decks
• An evaluation suite (evaluate) with both reference-based and reference-free metrics
High-Level Directory/Module Analysis
• SlidesLib/ – low-level Python API for image generation, plotting, PPTX creation, search, VQA
• generate/ – scripts to invoke an LLM to produce slide‐creation code (create_slide, create_slide_deck) plus prompt libraries and utilities
• refinement/ – scripts and prompts to refine first-pass code outputs
• scripts/ – code‐cleanup and execution harnesses for generated/refined Python programs
• autopresent/ – training & inference wrappers to fine-tune and run slide‐generation models on HuggingFace
• slidesbench/ – functions to build datasets from PPTX, generate reference code, plus a directory of example slide decks and train/test data
A. Data & Models
1. slidesbench Dataset (CSV + PPTX decks)
2. Pretrained Llama Models & Fine-tuned AutoPresent Weights (on HF Hub)
B. Core Library
• SlidesLib (image_gen, plotting, ppt_gen, etc.)
C. Generation Pipeline
1. generate/create_slide_deck.py (calls LLM via generate/utils.py)
2. Output: Python “programs” per slide
D. Refinement Pipeline
1. refinement/refine.py & refine_all.py (takes existing_slide + code + prompt)
2. Output: refined Python code
E. Execution Layer
• scripts/clean_generated_code.py → scripts/execute_each_file.py
• These ingest raw/refined code, call SlidesLib to produce .pptx or images
F. Training Pipeline
1. autopresent/train.py (loads slidesbench CSV, fine-tunes a model on HF)
2. autopresent/generate.py (uses fine-tuned model to produce code on test CSV)
G. Evaluation Suite
1. evaluate/slides_eval.py, evaluate/page_eval.py (reference-based)
2. evaluate/reference_free_eval.py (LLM-based design critique)
H. External Services & Tools
• OpenAI API (LLM inference)
• HuggingFace Hub (model weights)
• unoconv (pptx→jpg/pdf conversion)
Relationships & Data Flows
• slidesbench/train_data → autopresent/train.py → Fine-tuned Model → HF Hub
• Fine-tuned Model &/or Base Llama → generate scripts → Generated Code → refinement (optional) → scripts execution → SlidesLib calls → Slide assets (.pptx/.jpg)
• Slide assets + reference slides → evaluate modules → Metrics
• The core SlidesLib is invoked by both generated/refined code and by any user-written programs
Architectural Patterns & Principles
• Modular pipeline: separate stages for generation, refinement, execution, training, evaluation
• Library/API layer (SlidesLib) decoupled from orchestration scripts
• Use of external LLM inference and HF ecosystems for extensibility
• Reproducibility via slidesbench reference code generation and standardized CLI scripts