🚀 Qwen3-4B-I-1509

🧾 Model Overview

• 🏗️ Base Model: Qwen3-4B-Instruct-2507
• 🎯 Training Method: Reinforcement Learning (GRPO) with multiple reward functions

🏆 Reward Functions

1. 📝 Rule-based Reward
   ✔️ Checks correctness of function call name and arguments.
   ➕ Partial credit for matching subsets of arguments.
2. 🔒 Self-Certainty Reward
   ⚡ Encourages confident predictions.
3. 🔧 Tool-Call Reward
   ✅ Validates structural correctness.

⚙️ Training Configuration

• ⚡ Optimizer: AdamW
• 📉 Learning Rate: 5e-6 with cosine decay (min_lr_rate=0.1)
• ⏳ Scheduler: cosine_with_min_lr
• 🔄 Generations per Prompt: 4

📊 Eval Result:

Important notes:

• Why it lower than technical report?
  There have a limit of hardware so have to reduce some max tokens when evaluation for both 2 models
• Fair evaluate ?
  I use the same configuration for all the models I review for larger or with a same size model.

Tau-Bench

ACEBench

Contribute:

Support me at:

📖 Citation