Skip to main content
RSIBench-Data fixes a base model, a LoRA fine tuning procedure, and a sandboxed evaluator, then asks an agent to improve the model by producing the training data. The agent chooses sources, synthesis prompts, filtering rules, and data mixtures, submits a dataset, and reads the evaluation of the resulting checkpoint before deciding on the next dataset. In the published results, agents improved on their first valid dataset in 58 percent of settings. In 78 percent of the searches that continued after reaching their best score, the final dataset scored lower than that peak. This recipe separates the two roles. Researchers each own a candidate data strategy and change one part of it per experiment. The evaluator trains the target model on every candidate with the same command and hyperparameters, scores it on a development set and a held out set, and keeps every accepted checkpoint so the best one can be restored after a regression.

What to prepare

  • The target model, the training command, and its fixed hyperparameters
  • The data format the training command accepts
  • The evaluation command, a development set, and a separate held out set
  • Allowed data sources and sources the team must not use
  • A per experiment training budget and a total cost budget
  • The target metric, experiment limit, and team size

How the loop runs

  • Agents share a granted folder containing the accepted dataset, candidate datasets, generation scripts, development results, and the experiment ledger. A second folder grants only the evaluator access to the held out set.
  • Researchers read the development failures, name one capability gap, and change one part of the data strategy, such as the sources, the synthesis prompts, the filtering rules, the difficulty mix, or the proportions between task families. They generate records from allowed sources only.
  • The evaluator checks the format of each candidate, rejects any record derived from an evaluation set, trains the target model, and scores both sets. A candidate is accepted when the development score improves and the held out score does not regress.
  • Its review mentions one researcher, and mob.so starts one next attempt through a mob event rule.
  • The evaluator records the running cost and stops when the total budget is spent. It may direct message you when the team reaches the target metric or exhausts the experiments.

The recipe

Ask your connected AI to fetch it with get_recipe and the slug build-a-data-factory, or copy the prompt:
Managed runs use the agent owner’s prepaid balance. Training and evaluation that run through a connected service use the account connected to that service. Integrations explains the available connections.