Open agentic coding data (SFT + DPO), built forward, not distilled

Hi everyone,

We’re a small startup focused on agent data. We just released a dataset on the Hub, and we’d like to share the thinking behind it and get feedback from people who train or evaluate coding agents.

SFT: [logitly-ai/flash-coding-sft-v1 · Datasets at Hugging Face](https://huggingface.co/datasets/logitly-ai/flash-coding-sft-v1)

DPO: [logitly-ai/flash-coding-dpo-v1 · Datasets at Hugging Face](https://huggingface.co/datasets/logitly-ai/flash-coding-dpo-v1)

• SFT data: agentic coding trajectories for supervised fine-tuning
• DPO data: preference pairs for preference optimization
• Together, these can support offline reinforcement learning workflows

We build complex, controllable system environments and generate agentic coding data from them “forward”, meaning the data comes from real interaction with the environment. It is not distilled from proprietary models. We chose this for three reasons:

1. It scales. Once an environment is built, trajectory generation is reproducible and can be expanded in a controlled way. We think this is an important piece of infrastructure for future model progress.
2. Distillation has a short shelf life. Many providers’ terms of service restrict training on model outputs. Technically, access to reasoning traces is shrinking. If reasoning is streamed in encrypted form, or newer models don’t decode the thinking token, distillation-based improvement stops working. It can give a short-term ROI, but it is hard to build on.
3. Provenance matters. Data you generate yourself has a clear origin and can be improved in a principled way.

This dataset has no overlap with the original tasks in DeepSWE or Terminal-Bench. You can use it for training without worrying about contaminating those evaluations.