$ show aceliuchanghong/FAQ_Of_LLM_Interview
This is a personal interview preparation and learning repository focused on large language model (LLM) algorithm positions. It contains extensive interview question logs from 2024-2026, organized by company, covering topics such as fine-tuning, RAG, RLHF, agent design, model architecture, and deployment. The repository also includes learning notes on transformer, CNN, RNN, LoRA, PEFT, and other LLM-related concepts, along with code implementations: a decoder-only transformer trained on Chinese text (红楼梦), a reinforcement learning-based code optimizer using PPO, and various PyTorch examples. The utility scripts (DuckDB cache decorator, JSON extraction, logging, proxy management) are well-crafted and potentially reusable. However, the code components are explicitly unmaintained (the author states only interview records will receive future updates), and many paths are hardcoded. The interview notes are unstructured personal logs, not reusable as data assets without significant cleanup. The repo serves as a portfolio of technical capability and study material rather than a production system or reproducible data pipeline.
The LLM identified 4 files worth highlighting for reuse:
A reusable DuckDB-backed function cache decorator supporting sync/async, MD5 hashing for cache keys, pandas DataFrame serialization, and thread-safe locking. Demonstrates a clean, generic caching pattern that could be extracted and applied to any Python project needing persistent memoization with minimal dependency.
Robust JSON extraction from markdown strings, handling partial JSON, trailing commas, single quotes, and escaped characters. Includes a fallback to strict json.loads. This utility solves a common problem when extracting structured data from LLM outputs and is cleanly isolated for reuse.
A Gym environment for code optimization that applies AST-level transformations (e.g., for-loop to list comprehension, nested loop flattening) and evaluates reward based on runtime, brevity, and correctness. Demonstrates a non-obvious pattern for using reinforcement learning on code manipulation, with a clean interface.
A full implementation of a decoder-only transformer with ScaledDotProductAttention, multi-head attention, feed-forward network, and positional encoding. The code is modular, well-structured, and self-contained, making it a strong educational example of transformer internals. Easily extracted for learning or as a baseline.