shaffhausenquant/OpenThinkerAgent-32B-SFT-1K
OpenThinkerAgent-32B-SFT-1K is a 32 billion parameter instruction-tuned causal language model developed by OpenThoughts-Agent. Post-trained from Qwen3-32B, it is specifically fine-tuned for agentic tasks using the 1,000-example OpenThoughts-Agent-SFT-1K dataset. This model excels at agentic reasoning and task execution, demonstrating improved performance on benchmarks like OpenThoughts-TBLite and Terminal-Bench 2.0 compared to its base model.
Loading preview...
OpenThinkerAgent-32B-SFT-1K Overview
OpenThinkerAgent-32B-SFT-1K is a 32 billion parameter language model developed by the OpenThoughts-Agent team, specializing in agentic capabilities. It is a post-trained version of Qwen3-32B, fine-tuned using a full-parameter Supervised Fine-Tuning (SFT) approach on the OpenThoughts-Agent-SFT-1K dataset. This dataset comprises 1,000 high-quality (task, agent-trajectory) pairs, sourced from platforms like SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos, and IssueTasks, with trajectories generated by GLM-4.7-AWQ and filtered for traces with at least 5 model turns.
Key Capabilities & Performance
This model is designed to enhance agentic reasoning and task execution. Benchmarking in the terminus-2 harness shows notable improvements over its base model, Qwen3-32B, in specific agentic tasks:
- OpenThoughts-TBLite: Achieves 21.0 pass@1, outperforming Qwen3-32B's 13.7.
- Terminal-Bench 2.0: Scores 10.9 pass@1, compared to Qwen3-32B's 7.5.
Training involved a cutoff length of 32768 tokens, a learning rate of 4e-05, and 7 epochs, utilizing bf16 precision and DeepSpeed ZeRO-3.
Ideal Use Cases
OpenThinkerAgent-32B-SFT-1K is particularly well-suited for applications requiring advanced agentic behavior, such as:
- Automated problem-solving in technical domains.
- Complex task execution and multi-turn interactions.
- Development of AI agents for code generation, debugging, and system interaction.