Augustue/Qwen3-4B-Coder

Qwen3-4B Coder Task-Selected Checkpoint

This repository hosts a Qwen3-4B checkpoint selected as the code expert in our model merging experiments.

Although the checkpoint was selected from a broader pool of Qwen3-4B fine-tuned models, it shows strong code-generation ability under our evaluation protocol. With reasoning mode enabled, it achieves 89.02 pass@1 on HumanEval. This score is higher than several code-specialized checkpoints evaluated under the same protocol in our candidate pool.

We use this checkpoint as a task-selected code expert for model merging experiments, rather than assigning experts solely according to their original repository labels.

Evaluation

Benchmark	Setting	Score
HumanEval	reasoning mode, final submitted code evaluated	89.02 pass@1

Base Model

• Base model: Qwen/Qwen3-4B EOFcat > /tmp/README.md <<'EOF'

---

license: apache-2.0 base_model: Qwen/Qwen3-4B pipeline_tag: text-generation

Qwen3-4B Coder Task-Selected Checkpoint

This repository hosts a Qwen3-4B checkpoint selected as the code expert in our model merging experiments.

Although the checkpoint was selected from a broader pool of Qwen3-4B fine-tuned models, it shows strong code-generation ability under our evaluation protocol. With reasoning mode enabled, it achieves 89.02 pass@1 on HumanEval. This score is higher than several code-specialized checkpoints evaluated under the same protocol in our candidate pool.

We use this checkpoint as a task-selected code expert for model merging experiments, rather than assigning experts solely according to their original repository labels.

Evaluation

Benchmark	Setting	Score
HumanEval	reasoning mode, final submitted code evaluated	89.02 pass@1

Base Model

• Base model: Qwen/Qwen3-4B