<!-- agent-source-envelope -->
# posttrainllm — leaderboard

> Browser-trained, browser-runnable LLMs ranked on small benchmarks. Pareto views by params and by training compute.

Canonical: https://posttrainllm.com/leaderboard
<!-- /agent-source-envelope -->

[← playground](https://posttrainllm.com/playground)
[devlog](https://posttrainllm.com/devlog)
[roadmap](https://posttrainllm.com/roadmap)
[speedup](https://posttrainllm.com/speedup)

# Tiny models, browser-trained, ranked.

This is the **learning track** — small byte-level GPTs you can
train from scratch in a browser tab and run instantly. It is not
the factory's specialist evals (those live under
[specialists](https://posttrainllm.com/artifacts) and [experiments](https://posttrainllm.com/docs/attempt-ledger)). Each row is a `.tinygpt` checkpoint — hit **try →** to load it in the playground. Ranked by **TinyStories perplexity** (a 50-story holdout; lower is better); sort-6, reverse-16, and FineWeb tabs
below.

## Browser-trained benchmark results

FineWeb PPL (BPE)

[JSON](https://posttrainllm.com/data/leaderboard.json)
[CSV](https://posttrainllm.com/data/leaderboard.csv)

**TinyStories PPL.** Held-out perplexity over 50 stories from a slice of the TinyStories corpus that wasn't seen during gallery training. Score = exp(mean per-byte cross-entropy). Lower is better. The cleanest universal benchmark — every byte-level LM gets a real number, with clear differentiation between in-distribution and out-of-distribution models.

| # | Model | Params | Score | Train | Try |
| --- | --- | --- | --- | --- | --- |
| 1 | 📖 TinyStories · curated · browser · trained on roneneldan/TinyStories | 9.6M | 2.152 | 48 min | [try](https://posttrainllm.com/playground?gallery=tinystories) |
| 2 | 💬 Q&A chat · curated · browser · trained on databricks/databricks-dolly-15k | 9.6M | 5.346 | 50 min | [try](https://posttrainllm.com/playground?gallery=chat) |
| 3 | 🐟 Sea Tales · curated · browser · trained on Moby-Dick (first 800K) + Heart of Darkness — both public domain | 9.6M | 7.311 | 48 min | [try](https://posttrainllm.com/playground?gallery=seatales) |
| 4 | 🎭 Shakespeare · curated · browser · trained on tinyshakespeare (Karpathy) | 9.6M | 7.864 | — | [try](https://posttrainllm.com/playground?gallery=shakespeare) |
| 5 | ⌨️ Python code · curated · browser · trained on codeparrot/github-code-clean (Python) | 9.6M | 11.2 | 47 min | [try](https://posttrainllm.com/playground?gallery=code) |
| 6 | • flagship-huge · trained on BPE-trained (vocab 49152) | 34.6M | 85.9 | — | [try](https://posttrainllm.com/playground?gallery=flagship-huge) |

**Train one in your browser tab.** Pick a corpus, hit Start,
wait minutes-not-days. When the model converges, hit "Download model" — you
get a `.tinygpt` file. The gallery cards above are all browser-trained
9.6M-param models; a sort-6 / reverse-16 model can be 100× smaller and beat
them. Submission flow ships next — for now, open a PR adding your model to
`data/gallery/` and we'll score it.

[Train one →](https://posttrainllm.com/playground)
