| 📁 experiments/5_1_logprobs/ | Section 5.1: load a small model locally, read out every candidate token and its probability at each step, and see how temperature and top-p change sampling |
| 📁 experiments/5_2_base_vs_instruct/ | Sections 5.2 and 5.4: give the same prompts to a pretrained-only base model and a post-trained model and compare continuation, instruction following, sycophancy and style |
| 📁 experiments/5_5_thinking/ | Section 5.5: run one model with thinking on and off over a set of questions with known answers and compare accuracy, latency and cost |
| 📁 experiments/5_6_prefill_cache/ | Section 5.6: time the prefill and decode phases, and check whether a timestamp at the start of the prompt breaks the prefix cache (API and local llama.cpp versions) |
| 📁 experiments/5_7_local_speed/ | Section 5.7: measure single-stream and concurrent speed of a small model on your own machine and compare 4-bit and 8-bit quantization (MLX, llama.cpp, vLLM) |
| 📁 experiments/5_8_realtime/ | Section 5.8: first-word and final-transcript latency of real-time speech recognition, and whether a name list helps more during or after recognition |
| 📁 experiments/5_10_payback/ | Section 5.10: payback calculator for self-hosting, from your usage, GPU quotes, electricity and API prices |