Train your own 17-million-parameter text classifier in about a minute, with labels made by a big AI model, and know when it beats an API.

A 17M model (Ettin), fine-tuned in 42 seconds on one gaming GPU, sorts bank messages into 77 categories about as well as Claude Opus (91.5% vs 92%), for about $0 instead of $8,120 per million messages, at 5 ms a message. It also trains on a laptop CPU (11 min) and on a phone's GPU (21 min).

This repository has everything behind those numbers: the recipe, a step-by-step tutorial, the measured results on ten public tasks, the phone trainer, and the source of the explainer video.

• The model is the easy part; the labels are the whole game. More labelled examples keep helping, for every model size we tried (learning curves).
• No labels? Borrow a teacher. A big model (Kimi K3) labels your messages: 2,000 in about 20 seconds for about $1. The student trained on them lands a few points under its teacher; human labels break that ceiling.
• Ten tasks, one recipe. On human labels, Ettin 17M beats Jev (a zero-shot classification API) on 6 of 10 tasks. On Kimi's labels, it ties Jev on 2 and is within 5 points on 5 more; it loses where the teacher itself is weaker (table).
• When it pays: against Jev's price, fine-tuning pays for itself after roughly 50,000 to 500,000 messages (per task). Below that, or before you know you need scale, use an API.
• Evaluate first. To trust any model in production, you need a few hundred messages with checked answers. Those are also most of what you need to train your own.

The tutorial — tutorial/distill-ag-news.md: a runnable notebook, top to bottom in about 10 minutes and $1.50 of API credit. Tidy ten datasets, label AG News with two teachers (Kimi K3 and Qwen3.8-27B), fine-tune Ettin-17M on each set of labels and on the human labels, compare, and use the model. Our run: human labels 89.2%, Kimi's 85.1%, Qwen's 81.8%; 6 ms per message on a CPU.

The recipe, four scripts (uv installs everything on first run):

uv run recipe/prepare.py                         # download the 10 public datasets into data/tasks/<task>/
export OPENROUTER_API_KEY=...                    # and/or FIREWORKS_, TOGETHER_, DEEPINFRA_, PARASAIL_API_KEY
uv run recipe/label.py data/tasks/agnews         # Kimi K3 labels the 2,000 training messages (~20 s, ~$1)
uv run recipe/train.py data/tasks/agnews --labels labels/agnews-train.jsonl --save models/agnews
uv run recipe/train.py data/tasks/agnews         # the same on the human labels, for comparison
uv run recipe/breakeven.py                       # when fine-tuning pays, per task

from transformers import pipeline
classify = pipeline("text-classification", model="models/agnews", device="cpu")
classify("NASA's new telescope finds water vapour on a distant planet")   # [{'label': 'Sci/Tech', ...}]

Your own data: make a folder like data/tasks/agnews/: a task.json with {"name", "instruction", "labels"}, and train.jsonl / test.jsonl with one {"id", "text", "label"} per line (train.jsonl can leave label out when a teacher labels it). Check a few hundred test messages by hand: that is what every comparison stands on.

Ettin-17M is an open encoder from Johns Hopkins (a cousin of BERT, 17M parameters, most of them its vocabulary). It reads the whole message at once and gives one score per category. We add one output per category and train every weight: batch 32, AdamW with learning rate 1e-4 and weight decay 0.01, 5% warm-up then cosine decay, gradient clipping at 1.0, cross-entropy; 6 passes over ~9,500 messages, or the same number of steps for fewer (28 passes over 2,000). Nothing is tuned per task.

• results/: every number above, how it was measured, and the charts.
• phone/: the same training on a phone's GPU, in Chrome, with WebGPU and jax-js.
• video/: the source of the explainer video (script, animated charts, music and captions made from code; voice by Kokoro on a local GPU).

• Check the teacher's terms. Train only on answers from a model whose license and provider terms allow it. We used Kimi K3; the answers of some models and APIs may not be used to train other models.
• Numbers vary a little between runs (±0.5 points from the random seed; ±1.5–2 from test-set size). The tables report means of 3 runs.
• Datasets keep their own licenses; this repository downloads them, it does not redistribute them.
• NixOS: if training fails with /sbin/ldconfig not found, set TRITON_LIBCUDA_PATH=/run/opengl-driver/lib.

MIT license. By Maxime Rivest.