steer-v2 · RL & training
Short answer
1,229 of my own prompts to coding agents from my work in RL & training, Mar 2026 to Sep 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
What it is
Real prompts I wrote while steering coding agents on my own projects in this domain. Each record pairs the prompt with the agent's next actions (tool calls, where the harness recorded them) and my following turn, so the data shows how a human corrects, redirects and approves an agent over a long session.
Size and period
| Prompts | 1,229 |
| Sessions | 36 |
| Period | Mar 2026 to Sep 2026 |
| Harnesses | Codex (808), Pi (187), Claude Code (176), Codex (DeepSeek) (58) |
| With the agent's tool calls | 1,115 prompts |
| Projects it spans | 13 |
What it is good for
- Training and evaluating agents setting up SFT and RL runs on small models: data, rewards, evals and stability fixes.
- Studying how an experienced engineer steers agents: corrections, scope changes, approvals.
- Building reward models or graders for agent behaviour from real human follow-ups.
What is left out
- Model-written text of any kind: no agent replies, only my prompts and the agent's tool-call metadata.
- Client work, job applications and personal folders.
- Secrets and personal data: scrubbed by rules, then by a local privacy model as a second pass.
How to license it
Non-exclusive, per pack, directly with me or through DataVendor. Ask for the datasheet and a 25-record sample first.
Want this pack?
Ask for a sample →Tell me what you're training or evaluating.