Example compiles

This site does not store training runs. These ten files already compiled here. A real run lives on your machine, under your TINKER_API_KEY, after you download a loop and execute it.

Live catalog unavailable. You can still compile a loop and train on your machine.

How a run starts

Compile here. Train there.

1. Build a loop. 2. uv pip install tinker. 3. export TINKER_API_KEY=… 4. python your-loop.py. The file calls sample, forward_backward, optim_step, and save_state. Checkpoints belong to your Tinker project. Tinker docs: quickstart.

A page that listed your checkpoints would need your key in the browser or a login. That is not this product.

Open the builder

These ten walks already compiled. They are not live training jobs. They are the files you would run.

No.FileSignalTaskBase model
01tool-agents-nemotron-3-nano-environment.pyVerifiable outcomeSpecialized Tool AgentsNemotron-3-Nano
02formal-reasoning-deepseek-v3-1-environment.pyVerifiable outcomeFormal Reasoning EnginesDeepSeek-V3.1
03forecasting-qwen3-6-35b-a3b-environment.pyVerifiable outcomeCalibrated ForecastingQwen3.6-35B-A3B
04tool-agents-qwen3-8b-dpo.pyPreference pairsSpecialized Tool AgentsQwen3-8B
05forecasting-deepseek-v3-1-dpo.pyPreference pairsCalibrated ForecastingDeepSeek-V3.1
06formal-reasoning-qwen3-5-9b-dpo.pyPreference pairsFormal Reasoning EnginesQwen3.5-9B
07tool-agents-nemotron-3-nano-sdft.pyDemonstrationsSpecialized Tool AgentsNemotron-3-Nano
08forecasting-qwen3-5-4b-sdft.pyDemonstrationsCalibrated ForecastingQwen3.5-4B
09tool-agents-qwen3-8b-environment.pyVerifiable outcomeSpecialized Tool AgentsQwen3-8B
10formal-reasoning-qwen3-6-35b-a3b-sdft.pyDemonstrationsFormal Reasoning EnginesQwen3.6-35B-A3B
Example compiles — Reinforcement: Build Your Own Reward Model