AIBuildAI v2.5.9

AIBuildAI takes a dataset and a written description of the machine-learning task, has AI agents design and train several candidate models, scores them, and writes the predictions, a standalone inference script, and each candidate's trained files into the run's output directory.

This maintenance release continues the v2.5 line. Its four fixes are about a run keeping its candidates working instead of losing them or leaving the machine idle.

Fixes

• A run no longer dies with Segmentation fault when an agent's session ends in the middle of a turn. Before, that ended the run at once with no exit summary. Now the run stops with its exit summary and the line that tells you how to resume it.
• Candidates are now told how to pass the check that runs just before full training. Before, a candidate could die at that check and report an internal error instead of training; a candidate that follows the instruction now trains.
• After a candidate fails before training, the run looks for the next candidate within 30 seconds. Before, the freed slot could stay idle for up to 30 minutes with candidates waiting.
• Up to the parallel count you set now start together. Before, a run wrote and tested one candidate's code at a time, so a slow first candidate held the others back. Whenever the run declines to start another candidate, the log in the run directory says why.

Install

Download the tarball and SHA256SUMS from this page, then run:

:::bash
sha256sum -c SHA256SUMS
tar -xzf aibuildai-linux-x86_64-v2.5.9.tar.gz
cd aibuildai-linux-x86_64-v2.5.9
./install.sh
export PATH="$HOME/.local/bin:$PATH"
aibuildai --version

The installer puts the program under ~/.local, and adds ~/.local/bin to PATH in ~/.bashrc when it is not there already.

The tarball SHA256 is:

:::text
6b415ec066174351ced028c529879d92e38260d0963092fee714fa6ff9d10217  aibuildai-linux-x86_64-v2.5.9.tar.gz

Run a task

Sign in, generate the starter configuration, edit it, and start the run:

:::bash
aibuildai login
aibuildai config > task.yaml
aibuildai run task.yaml

In the starter configuration, replace the placeholder values of run.task_name, run.data_root, run.instruction, and run.playground_root. The instruction can be text or a file given as @/path/to/file.md. llm.model names a Claude model; keep it or replace it with the model you want. Review search.parallel and the budget fields before a first run: the starter starts six candidates at once and gives each candidate an hour-long training budget, which is more than a small machine or a first try needs.

For a Claude model, AIBuildAI uses an existing Claude Code sign-in on the machine. If there is none, run claude auth login or set AIBUILDAI_API_KEY to an Anthropic API key. For a bare deepseek-* model, set AIBUILDAI_API_KEY to a DeepSeek key. For another Anthropic-compatible endpoint, use a provider/model name and set AIBUILDAI_BASE_URL, AIBUILDAI_API_KEY, and AIBUILDAI_SMALL_FAST_MODEL. This model sign-in is separate from aibuildai login, which checks your AIBuildAI plan.

Use aibuildai whoami to check your AIBuildAI account and plan. Use aibuildai account to open the account page.

Requirements

• Linux x86_64 with glibc 2.29 or newer.
• 0.6 GB for the installed copy, 1.4 GB free during a first install, and 1.9 GB free during an upgrade. If the machine has no Conda, the first run installs one under your home directory. Task environments and run outputs need additional disk space.
• Network access for the model service and a writable output disk.
• Memory and compute capacity that fit your task. An NVIDIA GPU makes training faster; a machine without one runs on the CPU.
• A Max plan or above. aibuildai account opens the account page with the plans.