DEV Community

Cover image for AI companies pay billions for this data. I give it away for free.
Sumama-Jameel
Sumama-Jameel

Posted on

AI companies pay billions for this data. I give it away for free.

There is one thing AI companies want right now more than chips.

Data. Not any data. Recordings of coding that actually worked. Meta put 14 billion dollars into Scale AI. A company that sells training data. Not models. Data. And right now, Mercor pays engineers 80 to 200 dollars an hour to sit down and record real coding work for AI labs.

That is the food. It is real, it is rare, and it sits behind a paywall. I make that food. And I give it away. MIT license, Zero dollars, No signup.

Receipts: Meta's data deal and Mercor's own rates page.

What the food actually is

Not code. Not diffs. The whole work session.

A ticket. An agent reading a real repo. Tool calls. A test that fails. The fix. The retry. The pass. Every step in order. Every terminal output exactly as the tool said it. The AI world calls this a trajectory. I call it the struggle. A diff shows the answer. The struggle shows how someone got there. Models learn from the struggle.

Right now: 66 recordings. 2,541 real tool calls. Five real repos you already use: click, attrs, packaging, tenacity, cli.

"Free data is usually trash." Fair. Here is the difference.

  1. Everything ran for real, in a sandbox. No invented outputs. Ever.
  2. The answer was not in the room. I erase the git history, so the agent can not peek at the human fix.
  3. The tests are locked. The agent can not edit them to cheat.
  4. A recording ships only if the tests pass and the work stands next to the real human fix.
  5. The mistakes stay in. 60 failed runs, kept in order, with the fixes after them. Most sellers hide the mistakes. The mistakes are the lesson.
  6. You can check me. One command verifies all 66 files on your machine.

And this week I audited my own data. Found one bad episode. I did not delete it. I labeled it, locked it out of the training lists, and published the full report. Every flaw, in the open.

Show me a data seller who does that.

I said I would sell this. I changed my mind.

Some of you read my old post. I said this data was for sale. Here is why it is not.

The big labs trained on your code for free. Your repos, your posts, your answers. Then the cleaned up version got a price tag in the billions. I think the food should flow both ways. So the data is free. For the kid training a model on a free GPU. For the one person lab. For you.

And my selfish reason, straight up: a paywall with zero users is worth zero. I want people to train on it, break it, and tell me the numbers. That is worth more to me than money right now. The factory stays mine. The food is yours.

The numbers I promised, including the ugly ones

My last post promised real test numbers. Here they are. I trained a small Qwen model on this data. On my own test, correct tool calls went from 43% to 80%. Then the rest of my tests showed other scores got worse. But that was my engineering problem. Not a data issue.

I could have hidden that. Then I would be every other data seller. So it is in the README. I rebuilt the training pipeline and started retraining. No clean win yet.

And yes, 66 is small. One research team got a 47% jump on a tool use test from just 239 recordings. A little real data beats a mountain of fake data. And the factory does not sleep.

It is just there

I am not going to beg you to use it. It is just there. A door left open.

git clone https://github.com/Sumama-Jameel/megabit-trajectories.git
cd megabit-trajectories
python3 verify.py
Enter fullscreen mode Exit fullscreen mode

That last line checks all 66 recordings. If it says 66/66, you have exactly what I published. Byte for byte. It ships in OpenAI chat format, with converters for other formats included.

Repo: https://github.com/Sumama-Jameel/megabit-trajectories

Train on it. Report the numbers. Even the bad ones. Especially the bad ones.

One last thing

Last time I opened a free door to 75 models. Models eat data. This is the free food. They trained on your code for free. Now training data sells for billions.

Should this food be free for everyone, or is selling it fair? Pick a side. I will be down there.

Top comments (0)