{"title":"Exercise: Understanding LLM Parameters","description":null,"content":"# Exercise 1: Sweep the Temperature\n\nSend the same prompt at three different temperatures and compare the outputs, so you can see how temperature controls the randomness of a model's response.\n\nTemperature is the parameter people reach for first and understand least. The fastest way to build intuition for it is to hold everything else fixed and watch what changes as you turn it up. Low temperature gives focused, repeatable answers; high temperature gives varied, surprising ones.\n\n### Requirements\n\n- Use one fixed prompt that invites some creativity, such as naming a product.\n- Call the model at temperatures `0.0`, `1.0`, and `2.0`.\n- Run each temperature a few times so you can see how much the output varies.\n- Print the temperature alongside each response.\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport os\nfrom dotenv import load_dotenv\n\nload_dotenv()\n\nclient = OpenAI(\n api_key=os.getenv(\"OPENROUTER_API_KEY\"),\n base_url=\"https://openrouter.ai/api/v1\",\n)\n\nprompt = \"Suggest a single creative name for a developer productivity app. Reply with just the name.\"\n\n# TODO: for each temperature in [0.0, 1.0, 2.0]:\n# call the model 3 times with that temperature\n# print the temperature and each response\n```\n\n\n### Expected Output\n\nAt temperature `0`, the three runs are nearly identical. As temperature rises, the runs diverge.\n\n\n```plaintext\nTemperature 0.0:\n ShipFast\n ShipFast\n ShipFast\nTemperature 1.0:\n FlowForge\n DevPulse\n CodeCadence\nTemperature 2.0:\n Zephyrline\n Quillstack\n Nimbello\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport os\nfrom dotenv import load_dotenv\n\nload_dotenv()\n\nclient = OpenAI(\n api_key=os.getenv(\"OPENROUTER_API_KEY\"),\n base_url=\"https://openrouter.ai/api/v1\",\n)\n\nprompt = \"Suggest a single creative name for a developer productivity app. Reply with just the name.\"\n\nfor temperature in [0.0, 1.0, 2.0]:\n print(f\"Temperature {temperature}:\")\n for _ in range(3):\n response = client.chat.completions.create(\n model=\"openai/gpt-5.4-mini\",\n messages=[{\"role\": \"user\", \"content\": prompt}],\n temperature=temperature,\n )\n print(f\" {response.choices[0].message.content}\")\n```\n\n\nHolding the prompt and model fixed isolates temperature as the only variable. The inner loop runs each setting three times so the variation is visible rather than something you have to infer from one sample. The contrast between the repeatable output at `0` and the scattered output at `2.0` is the whole point: temperature is the dial you set low when you need consistency and higher when you want variety.\n\n---\n\n# Exercise 2: Control Length with max_tokens and stop\n\nTemperature controls randomness; `max_tokens` and `stop` control where generation ends. Use both on the same prompt and read the `finish_reason` to see why each response stopped: hitting the token cap versus reaching a stop sequence.\n\n### Requirements\n\n- Call the model with a small `max_tokens` and print the content and `finish_reason`.\n- Call again with a `stop` sequence and print the content and `finish_reason`.\n- Compare how each response was cut off.\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport os\nfrom dotenv import load_dotenv\n\nload_dotenv()\n\nclient = OpenAI(\n api_key=os.getenv(\"OPENROUTER_API_KEY\"),\n base_url=\"https://openrouter.ai/api/v1\",\n)\n\nprompt = \"List the eight planets of the solar system, one per line, in order from the sun.\"\n\n# TODO: call once with a small max_tokens (e.g. 16) and print content + finish_reason\n\n# TODO: call again with stop=[\"Mars\"] and print content + finish_reason\n```\n\n\n### Expected Output\n\nThe capped run stops mid-list with finish_reason 'length'; the stop run halts right before the stop sequence with finish_reason 'stop'.\n\n\n```plaintext\nmax_tokens run, finish_reason = length\nMercury\nVenus\nEarth\nstop run, finish_reason = stop\nMercury\nVenus\nEarth\n\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport os\nfrom dotenv import load_dotenv\n\nload_dotenv()\n\nclient = OpenAI(\n api_key=os.getenv(\"OPENROUTER_API_KEY\"),\n base_url=\"https://openrouter.ai/api/v1\",\n)\n\nprompt = \"List the eight planets of the solar system, one per line, in order from the sun.\"\n\n# max_tokens cuts the response off; finish_reason becomes \"length\"\ncapped = client.chat.completions.create(\n model=\"openai/gpt-5.4-mini\",\n messages=[{\"role\": \"user\", \"content\": prompt}],\n max_tokens=16,\n)\nprint(\"max_tokens run, finish_reason =\", capped.choices[0].finish_reason)\nprint(capped.choices[0].message.content)\n\n# stop halts generation when the sequence appears; finish_reason becomes \"stop\"\nstopped = client.chat.completions.create(\n model=\"openai/gpt-5.4-mini\",\n messages=[{\"role\": \"user\", \"content\": prompt}],\n stop=[\"Mars\"],\n)\nprint(\"stop run, finish_reason =\", stopped.choices[0].finish_reason)\nprint(stopped.choices[0].message.content)\n```\n\n\n`max_tokens` is a hard ceiling on output: when it is reached the response is truncated and `finish_reason` is `length`, which is the signal to raise the limit or shorten the task. A `stop` sequence ends generation as soon as that text would appear, giving `finish_reason` `stop`. Reading `finish_reason` rather than guessing is how you tell a complete answer from a truncated one, which matters when you parse or store the output.","pageType":"ai-engineering"}

Get Premium