Skip to content
ppile

Meet pile.

Documents in. Training data out.

Any shape, any mess. Each one enters sealed.

Get startedSee the studio
macOSLinuxCLI + studio
pile studio: drop documents, feed lineage, sealed ledger

Local on your machine. Corpus to training JSONL. Nothing leaves unless you export.

Drop documents. Each one enters sealed.

Any shape, any mess. Point pile at a folder or drop files into the studio. Readable copies plus training JSONL, every admission sealed with its cause.

Reading it as: corpus to training JSONL. Or ask the record a question.

Get started
pile drop zone and Run button

Lineage that names choosers

Feed shows who chose each document: a person, a default walk, or an agent. Orphans surface by name. Pressing Run never launders a corpus into being someone's choice.

Hashes, section counts, training records. The shape of the corpus, not the contents.

Feed lineage with chooser metadata

Workflows you can seal

documents.compile converts sources into readable artifacts and a training dataset. Every admission is sealed with its cause.

Save and seal. Starts by hand, on drop, on a rhythm, or on a timetable.

Workflow editor with documents.compile and Save and seal

Train where Unsloth already runs

Unsloth Studio connected on localhost. Pick a sealed run, set the knobs, keep or curtail the dataset.

Pile is everything before the GPU. Unsloth is the compute.

See integrations
Training tab with Unsloth Studio connected and knobs

Stranger verification

Status bar: Sealed · post quantum. Hand someone a chain file. No account, no key exchange, no trust in you.

Ed25519 and ML-DSA-65. The corpus stays private; the shape is what gets attested.

See the commands
Sealed post-quantum status in pile

Train where you already train

PILE sits between documents and trainers. Unsloth is the compute; pile is everything before.

View setup →
Unsloth

Pipe JSONL into Studio

Same artifacts your fine-tune expects, plus a provenance sidecar.

axolotl

Standard training exports

Parsers stay swappable underneath. Trainers stay interchangeable on top.

RunPod

Same corpus, your GPU

Cloud routing with transparent rates, after the sealed corpus is built.

MCP

Corpus tool for agents

bin/pile mcp. chooser=agent sealed like a click.

pile

Your complete training-data workspace

Run the layer before the GPU on Mac and Linux. Compile, seal, and export JSONL your trainer already reads, while the corpus stays private.

  1. 01

    Drop documents. Compile corpus to training JSONL in minutes.

  2. 02

    See who chose every document before you train.

  3. 03

    Hand a regulator a sealed chain: post quantum, stranger-verifiable.

Pricing

Priced like a professional tool

Local-first buyers get a one-time option. Subscriptions stay available if you prefer them. Cloud compute, when it ships, bills separately.

Trial

Free

Full features, capped at a small number of documents

Start trial

Personal

Popular

$149

One time, or $19 a month. One machine, individual use.

Get Personal

Commercial

$499

One time, or $49 a month. Company use, several machines.

Get Commercial

Local-first · sealed · post quantum

Ready to compile your corpus?

Design partners: drop documents, seal admissions, pipe JSONL into Unsloth.

Get startedBook a demo

Zero verb

bin/pile demo/corpus
bin/pile orphans
bin/pile disclose
bin/pile ledger verify