Multiturn-RL

Multiturn RL experiments

Setup

$ uv sync

Run

$ cd experiments
$ ./train.sh

TODO

• Make custom chat scheduler for terminal environment
• Create custom reward functions for various tasks (SWE-Bench, TACO, math, etc.)
• Collect data
  • ToRL dataset?
• Run experiments

Data

*1. Math

• NuminaMath
• AIME
• MATH
• etc.

2. Coding
   • TACO *3. SWE
   • SWE-Bench

Notes:

• should we do regular RL training before integrating with terminal? seems like it will take a while before the model starts using the terminal