| import altair as alt | |
| import numpy as np | |
| import pandas as pd | |
| import streamlit as st | |
| import streamlit as st | |
| st.set_page_config(page_title="Post-Training Techniques for LLMs", layout="centered") | |
| st.title("🔧 Beyond Pretraining: Post-Training Techniques for LLMs") | |
| st.subheader("Distillation, Preference Optimization, and RLHF — Visualized") | |
| st.markdown(""" | |
| This Streamlit app hosts a visual guide to help navigate post-training strategies for language models, with real-world examples like **SmolLM3**, **Tulu**, and **DeepSeek-R1**. | |
| 📎 Download the full slide deck: | |
| 👉 [Click here to download (PDF)](https://huggingface.co/spaces/bird-of-paradise/post-training-techniques-guide/blob/main/src/Post%20Training%20Techniques.pdf) | |
| --- | |
| 🧠 **Topics covered:** | |
| - Tradeoffs between SFT, DPO/APO/GRPO, PPO | |
| - Subjective vs Rule-based rewards | |
| - How real open-source models chose their strategy | |
| Made with ❤️ by Jen Wei | |
| """) | |
| # Optional: Slide preview | |
| st.image("src/Post_Training_Techniques_preview_2.png", caption="Slide 1: Tradeoffs between Optimization Paths", use_container_width=True) | |