liangzhidanta/Qwen3-8B-CC-SFT-v1

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwen3-8B-CC-SFT-v1 by liangzhidanta is an 8 billion parameter coding-agent SFT checkpoint based on Qwen/Qwen3-8B, fine-tuned with execution-verified GLM-5.3 teacher trajectories. This model is specifically designed for software engineering agents and tool-use research, serving as an initialization for subsequent verifiable coding-agent RL experiments. It achieves a Pass@1 score of 34.98% on the internal canonical303 evaluation, significantly outperforming its base model.

Loading preview...

Overview

Qwen3-8B-CC-SFT-v1 is an 8 billion parameter coding-agent SFT (Supervised Fine-Tuning) checkpoint developed by liangzhidanta. It is built upon the Qwen/Qwen3-8B base model and has been fine-tuned using execution-verified GLM-5.3 teacher trajectories. This model represents the first stage of agent SFT, specifically targeting software engineering tasks and serving as an initialization for further reinforcement learning (RL) experiments in verifiable coding agents.

Key Capabilities

  • Enhanced Coding Agent Performance: Achieves a Pass@1 score of 34.98% on the internal canonical303 evaluation, a substantial improvement over the base Qwen3-8B's 3.30%.
  • Full-Parameter Fine-Tuning: All 8.19 billion parameters were updated during SFT, without the use of LoRA, PEFT, or quantization.
  • Teacher-Guided Learning: Trained on 1003 execution-verified multi-turn trajectories generated by GLM-5.3, focusing on SWE-smith software-engineering tasks.
  • High Context Handling: Supports a training context of 32768 tokens and is evaluated with a 40960-token context using Claude Code native automatic context compaction.

Good For

  • Coding-Agent Research: Ideal for exploring and developing advanced software-engineering agents.
  • Tool-Use Research: Suitable for experiments involving language models interacting with external tools.
  • Multi-Turn Agent SFT/RL Initialization: Serves as a strong warm-start model for subsequent verifiable coding-agent reinforcement learning experiments.
  • SWE-Style Task Experimentation: Useful for tackling and experimenting with software engineering-style problems.

Finetunes
1 models