arvindcr4/tinker-rl-bench-scale_gsm8k_qwen3.5-4b

scale_gsm8k_qwen3.5-4b

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

Training Details

• Base model: Qwen/Qwen3.5-4B
• 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: 96.2%
• Last-10 avg reward: 85.0%
• Peak reward: 100.0%
• Zero-loss steps: 57%
• Tinker Run ID: 200e4983-3e9d-5176-87d0-c55dd3c73539:train:0

Reward Trace

[
  1.0,
  0.8125,
  1.0,
  1.0,
  1.0,
  0.5,
  0.5,
  0.875,
  0.625,
  0.8125,
  1.0,
  1.0,
  1.0,
  0.625,
  0.875,
  0.25,
  1.0,
  0.875,
  0.6875,
  0.5625,
  0.9375,
  0.5,
  0.9375,
  0.5,
  1.0,
  1.0,
  0.625,
  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}
}