arvindcr4/tinker-rl-bench-cross_tool_qwen3-32b

cross_tool_qwen3-32b

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

Training Details

• Base model: Qwen/Qwen3-32B
• Method: GRPO (Group Relative Policy Optimization)
• Platform: Tinker API v0.18.1
• Task: tool_use
• Seed: 42
• LoRA rank: 32
• Learning rate: 3e-05
• Group size: 8
• Steps: 30

Results

• First-5 avg reward: 0.0%
• Last-10 avg reward: 0.0%
• Peak reward: 0.0%
• Zero-loss steps: 100%
• Tinker Run ID: 73ff186a-d8e3-50c3-afb8-2b863cd09579:train:0

Reward Trace

[
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0,
  0.0
]

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}
}