arvindcr4/tinker-rl-bench-arch_gsm8k_kimi-k2

arch_gsm8k_kimi-k2

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

Training Details

• Base model: moonshotai/Kimi-K2-Thinking
• Method: GRPO (Group Relative Policy Optimization)
• Platform: Tinker API v0.18.1
• Task: gsm8k
• Seed: 42
• LoRA rank: 16
• Learning rate: 1e-05
• Group size: 4
• Steps: 20

Results

• First-5 avg reward: 92.5%
• Last-10 avg reward: 80.0%
• Peak reward: 100.0%
• Zero-loss steps: 40%
• Tinker Run ID: 51a8ef9e-15ef-5f8f-bda1-78ee51387a12:train:0

Reward Trace

[
  1.0,
  0.875,
  1.0,
  1.0,
  0.75,
  1.0,
  0.75,
  1.0,
  0.75,
  0.875,
  1.0,
  1.0,
  1.0,
  0.5,
  0.875,
  0.375,
  0.875,
  0.75,
  0.875,
  0.75
]

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}
}