liangzhidanta/Qwen3-8B-CC-SFT-v2-2e
Qwen3-8B-CC-SFT-v2-2e is an 8 billion parameter Qwen3-based causal language model developed by liangzhidanta, specifically fine-tuned for coding agent tasks. This model is the second epoch of compact-aware continued SFT, demonstrating enhanced performance on the canonical303 coding-agent benchmark. It excels at handling complex coding challenges within a 32K context window, making it suitable for automated code generation and problem-solving.
Loading preview...
Model Overview
Qwen3-8B-CC-SFT-v2-2e is an 8 billion parameter model from the Qwen3 family, developed by liangzhidanta. It represents the second epoch of a compact-aware continued Supervised Fine-Tuning (SFT) process, building upon the Qwen3-8B-CC-SFT-v2 checkpoint. This model is specifically optimized for coding agent tasks and demonstrates improved performance on relevant benchmarks.
Key Capabilities & Performance
- Enhanced Coding Agent Performance: Achieves 54.79% Pass@1 on the canonical303 coding-agent benchmark, solving 166 out of 303 tasks. This is a significant improvement over its predecessor (Qwen3-8B-CC-SFT-v2 at 47.52%).
- Compact-Aware Fine-Tuning: Trained on a specialized dataset, liangzhidanta/claude-code-glm53-swesmith-compact-trajectories, which includes execution-verified agent trajectories and segments derived from Claude Code's native compact requests.
- Extended Context Handling: Supports a context length of 40960 tokens, with context compaction enabled during evaluation, making it suitable for complex, multi-turn coding interactions.
- Robust Training: Underwent full-parameter continued SFT over two epochs, utilizing a fresh optimizer and cosine learning rate schedule for the second epoch.
Ideal Use Cases
- Automated Code Generation: Best suited for applications requiring an AI agent to generate and solve coding problems.
- Code Debugging and Refactoring: Its compact-aware training makes it effective for tasks involving iterative code improvements and understanding compacted context.
- Developer Tooling: Can be integrated into development environments for intelligent code assistance and agent-driven development workflows.