arvindcr4/tinker-rl-bench-arch_gsm8k_gpt-oss-20b

arch_gsm8k_gpt-oss-20b

GRPO experiment from TinkerRL-Bench world-class experiment suite.

Training Details

• Base model: openai/gpt-oss-20b
• Method: GRPO (Group Relative Policy Optimization)
• Platform: Tinker API v0.18.1
• Task: gsm8k
• Seed: 42
• LoRA rank: 32
• Learning rate: 3e-05
• Group size: 8
• Steps: 30

Results

• First-5 avg reward: 87.5%
• Last-10 avg reward: 87.5%
• Peak reward: 100.0%
• Zero-loss steps: 53%
• Tinker Run ID: c465d654-a00f-5821-9c11-0e40f0eecf9f:train:0

Reward Trace

[
  0.625,
  0.8125,
  1.0,
  0.9375,
  1.0,
  0.5625,
  0.6875,
  1.0,
  0.5,
  0.9375,
  1.0,
  0.9375,
  1.0,
  0.5,
  0.9375,
  0.5625,
  1.0,
  0.8125,
  0.625,
  1.0,
  1.0,
  0.5,
  0.9375,
  0.5,
  1.0,
  1.0,
  0.8125,
  1.0,
  1.0,
  1.0
]

Citation

@misc{tinker-rl-bench-2026,
  title={TinkerRL-Bench: A Unified Benchmark for RL Post-Training},
  author={Arvind C R and Sandhya Jeyaraj and Madhu Kumara L and Mohammad Rafi and Dhruva N Murthy and Arumugam K},
  year={2026},
  url={https://github.com/arvindcr4/tinker-rl-lab}
}