Start with a question.
Keep going in code.

01 / FIRST PREDICTION

Use the playground.

Begin with What is Pebble?, What is attention?, Explain tensors. or Explain functions..

1. Enter a short question using 1–160 ASCII characters.
2. Keep temperature at 0 for greedy generation with seed 2026.
3. Select Generate answer. CPU inference may take a few seconds.
4. Read the result, compare it with the recorded examples, and use Copy output if you want to keep it.

The web demo generates up to 96 character tokens and may stop earlier at EOS. Higher temperature, up to 1, adds randomness; it doesn't increase accuracy. Inference runs on the server using the released checkpoint, not on your device.

Open Reproducible sampling settings to choose a whole-number seed from 0 to 2147483647 and an output limit from 16 to 96 characters. Identical prompt, temperature, seed and output limit reproduce a sample. The cancel button cancels your browser request; the server may finish work already started. Your last three successful experiments stay in page memory only.

02 / SET EXPECTATIONS

A small model.
A specific task.

PebbleLM learns short answers from a forty-topic synthetic curriculum. It isn't trained for general chat, reliable coding advice, new knowledge or conversation memory.

Topics and target answers overlap across splits. The results test phrasing, not broad intelligence. Fluent output can still describe the wrong topic.

• 101-token vocabulary: printable ASCII, newline, tab and four special tokens. The web demo rejects non-ASCII; the CLI maps unsupported characters to UNK.
• 299-position context: character tokens, not words. This does not imply long-context capability.
• No retrieval or tools: there is no search, chat history or instruction hierarchy.

03 / YOUR MACHINE, YOUR WEIGHTS

Run it on your CPU.

Install Node 24, Git and npm. You don't need a GPU, Python trainer, external model provider or API key.

git clone https://github.com/YashAnand69/pebble-llm.git
cd pebble-llm
npm ci
npm run verify:artifacts
npm run inspect
npm run generate -- --prompt "What is Pebble?" --temperature 0 --max-tokens 96 --seed 2026

inspect confirms two million parameters and the WASM backend. verify:artifacts checks the released file hashes. The checkpoint is approximately 8 MB; runtime memory also includes tensor buffers, activations and the interpreter.

Explore sampling

npm run generate -- --prompt "What is attention?" --temperature 0.5 --top-k 10 --max-tokens 96

The CLI also supports --checkpoint and --seed. Its --max-tokens range is 1–1000. It loads checkpoints/best.pebble-weights by default. A longer output limit does not fix model quality.

Run the website locally

npm run build
npx vercel link
npx vercel dev

Choose your Vercel project when linking. Serving public/ alone previews the page but does not run inference. Vercel account access is needed for this workflow; no application environment variables are required.

04 / A SMALL JSON INTERFACE

Call the demo API.

Send a POST request from your terminal or server-side code. No model API key is needed.

curl https://pebble-llm.vercel.app/api/generate \
  -H 'Content-Type: application/json' \
  --data '{"prompt":"What is Pebble?","temperature":0,"seed":2026,"maxNewTokens":96}'

A successful response includes answer, parameters (2000000), elapsedMs and the actual seed, temperature and maxNewTokens. Elapsed time measures the model run and varies by question and server conditions.

Node 24 / server-side JavaScript

const response = await fetch(
  'https://pebble-llm.vercel.app/api/generate', {
    method: 'POST',
    headers: {'Content-Type': 'application/json'},
    body: JSON.stringify({
      prompt: 'What is Pebble?', temperature: 0,
      seed: 2026, maxNewTokens: 96
    })
  }
);
const result = await response.json();
if (!response.ok) throw new Error(result.error);
console.log(result.answer);

Use the full URL from server-side code. Same-origin browser code can use /api/generate; cross-origin browser requests are not supported by the API's CORS policy.

Invalid input returns 400. Unsupported methods return 405 with Allow: POST. Failed inference returns a generic 500. Platform timeouts may return non-JSON responses, so handle those in real integrations.

05 / READ, CHANGE, REPEAT

Retrain and evaluate.

Start with model.pebble, tokenizer.pebble and sampling.pebble. The architecture uses four decoder blocks, four heads, width 200, causal attention, pre-LayerNorm and tied embeddings.

npm test
npm run prepare:data
npm run train -- --steps 2000 --batch 2 --length 299 --seed 2026 --lr 0.001 --report-every 100
npm run refine -- --steps 1000 --batch 4 --seed 2027 --lr 0.0005 --report-every 100
npm run evaluate
npm run evaluate:canonical

Validation selects the checkpoint; keep reserved prompts out of training and selection. The published main run took 43.27 minutes and refinement took 12.16 minutes on an Apple M5 Pro. Your machine may differ.

06 / WHEN SOMETHING LOOKS WRONG

Troubleshooting.

Code, original data and released weights are MIT licensed. TensorFlow.js/WASM dependencies retain their Apache-2.0 licenses.