🛝LLM Playground🎈

What this is: Turns your text into the actual tokens a model sees — not words, but sub-word pieces from GPT-2's vocabulary.

Try this first: Type your own name, then try a word like 'tokenization' itself.

What to notice: Common words become one token; rare or made-up words fragment into pieces. Try a non-English sentence and watch it fragment even more — tokenizers are trained mostly on English text, and everyone else pays for it in extra tokens.

What this is: Generates text from GPT-2 with adjustable decoding settings — temperature, top-k, top-p, and a sampling on/off toggle.

Try this first: Turn 'do_sample' off and generate something. Then turn it on with temperature around 0.7.

What to notice: With sampling off, the model often loops — repeating the same sentence over and over. That's not a broken model, it's greedy decoding always picking the single most likely next word. Sampling breaks the loop.

What this is: Runs the same prompt through three decoding strategies at once — greedy, balanced, and high-temperature — so you can see the difference side by side.

Try this first: Use any prompt you like and compare all three outputs.

What to notice: Same model, same prompt, three very different personalities. Decoding strategy shapes the output as much as the model itself does.

What this is: Runs the same prompt across up to four different models — a base (pre-trained only) model and instruction-tuned ('chat') models — side by side.

Try this first: Compare 'qwen2.5-0.5b-base' against 'qwen2.5-0.5b-instruct' — same model family, only one has gone through post-training.

What to notice: The instruct model answers directly and formats well; the base model just continues the text. This is what post-training (SFT/RLHF) actually buys you — though watch closely, because well-formatted answers aren't always correct answers.

What this is: Every experiment you run gets logged automatically. This tab charts and lists them — repetition rate, perplexity, and judge scores across models and settings.

Try this first: Run a few generations first, then check this tab to see them charted.

What to notice: Use 'Group by Experiment' to see all results from one comparison together, and check the perplexity chart's caution note — it's only valid comparing a model against itself, not across different model families.

What this is: Submit a prompt and response to be scored by three AI judges of different sizes — a small local model, a larger local model, and a frontier API model — against the same rubric.

Try this first: Pick a past result from history, or write your own. Try something you know is subtly wrong.

What to notice: Watch whether the small judge actually catches the problem, or scores it well anyway. This surfaced some of the most interesting findings in the whole project — smaller judges tend to score generously regardless of actual quality.

What this is: A real multi-turn conversation interface with a chat-tuned model. Only instruction-tuned models can hold a conversation — base models don't have this tab available to them for a reason explained in the blog.

Try this first: Tell it something in one message, then ask about it two messages later.

What to notice: The token counter shows the conversation growing — models have no memory of their own; the whole conversation gets resent every single turn.