Adapters from the SDK course and community recipes, with the eval rows beside the weights. Read the recipe before the number.
AI & ML interests
Agents make mistakes. Those mistakes are the most valuable training data there is. Only While. turns them into a dataset, post-trains an open model on it with SFT and RL, and proves the agent stopped repeating them on a held-out test.
Recent Activity
The front door. These also appear under their use case below.
Training how an agent talks without changing what it knows. Judge scores the manner, code checks no required fact was dropped.
Constitutions and identity. A judge reads the reply against the principle, because code cannot grade manner.
Payment-intent data and the small models that verify what a shopper asked for before an agent acts on it.
Post-training papers rerun on small open models: both arms, one protocol, paired deltas. Recipe arm at the root, baseline in a subfolder.
Two distinct failures with different causes and different fixes: jailbreaks arrive from the user, prompt injections arrive inside a tool result.
Text-to-SQL against a live warehouse. Graded by executing the query, not by reading it.
The same task at the same quality with fewer tool calls.
Adapters from the SDK course and community recipes, with the eval rows beside the weights. Read the recipe before the number.
Post-training papers rerun on small open models: both arms, one protocol, paired deltas. Recipe arm at the root, baseline in a subfolder.
The front door. These also appear under their use case below.
Two distinct failures with different causes and different fixes: jailbreaks arrive from the user, prompt injections arrive inside a tool result.
Training how an agent talks without changing what it knows. Judge scores the manner, code checks no required fact was dropped.
Text-to-SQL against a live warehouse. Graded by executing the query, not by reading it.
Constitutions and identity. A judge reads the reply against the principle, because code cannot grade manner.
The same task at the same quality with fewer tool calls.
Payment-intent data and the small models that verify what a shopper asked for before an agent acts on it.