Parameter	Value
Base model	`Qwen/Qwen3-8B`
Method	GRPO (Group Relative Policy Optimization)
Environment	`tool_use`
LoRA rank	32
Learning rate	3e-05
Steps	2
Group size	8
Final avg reward	0.902
Tinker run ID	`eab60069-db43-5f60-bc58-4164bf576056:train:0`
W&B	https://wandb.ai/arvindcr4-pes-university/skyrl-tinker/runs/68h67x6y

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B")
model = PeftModel.from_pretrained(base, "arvindcr4/skyrl-tinker-qwen3-8b-tool_use")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

Training Curve

Trained for 2 steps with final average reward of 0.902.

Framework

RL Framework: SkyRL
Training Backend: Tinker (hosted)
Algorithm: GRPO